Post on 19-Jul-2020
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
MUTUAL FUND PERFORMANCE AND FLOWSLuboš Pástor and M. Blair Vorsatz
FIRM-SURVEY EVIDENCE IN EMERGING MARKETSThorsten Beck, Burton Flynn and Mikael Homanen
FINANCIAL MARKETS AND NEWSHarry Mamaysky
RACE AND REDLININGGraziella Bertocchi and Arcangelo Dimico
AGE AND INCOME INEQUITIESMichèle Belot, Syngjoo Choi, Egon Tripodi, Eline van den Broek-Altenburg, Julian C. Jamison and Nicholas W. Papageorge
CONSUMPTION AND GEOGRAPHIC MOBILITYRaymundo M. Campos-Vazquez and Gerardo Esquivel
ISSUE 38 16 JULY 2020
Covid Economics Vetted and Real-Time PapersCovid Economics, Vetted and Real-Time Papers, from CEPR, brings together formal investigations on the economic issues emanating from the Covid outbreak, based on explicit theory and/or empirical evidence, to improve the knowledge base.
Founder: Beatrice Weder di Mauro, President of CEPREditor: Charles Wyplosz, Graduate Institute Geneva and CEPR
Contact: Submissions should be made at https://portal.cepr.org/call-papers-covid-economics. Other queries should be sent to covidecon@cepr.org.
Copyright for the papers appearing in this issue of Covid Economics: Vetted and Real-Time Papers is held by the individual authors.
The Centre for Economic Policy Research (CEPR)
The Centre for Economic Policy Research (CEPR) is a network of over 1,500 research economists based mostly in European universities. The Centre’s goal is twofold: to promote world-class research, and to get the policy-relevant results into the hands of key decision-makers. CEPR’s guiding principle is ‘Research excellence with policy relevance’. A registered charity since it was founded in 1983, CEPR is independent of all public and private interest groups. It takes no institutional stand on economic policy matters and its core funding comes from its Institutional Members and sales of publications. Because it draws on such a large network of researchers, its output reflects a broad spectrum of individual viewpoints as well as perspectives drawn from civil society. CEPR research may include views on policy, but the Trustees of the Centre do not give prior review to its publications. The opinions expressed in this report are those of the authors and not those of CEPR.
Chair of the Board Sir Charlie BeanFounder and Honorary President Richard PortesPresident Beatrice Weder di MauroVice Presidents Maristella Botticini Ugo Panizza Philippe Martin Hélène ReyChief Executive Officer Tessa Ogden
Editorial BoardBeatrice Weder di Mauro, CEPRCharles Wyplosz, Graduate Institute Geneva and CEPRViral V. Acharya, Stern School of Business, NYU and CEPRAbi Adams-Prassl, University of Oxford and CEPRGuido Alfani, Bocconi University and CEPRFranklin Allen, Imperial College Business School and CEPRMichele Belot, European University Institute and CEPRDavid Bloom, Harvard T.H. Chan School of Public HealthNick Bloom, Stanford University and CEPRTito Boeri, Bocconi University and CEPRAlison Booth, University of Essex and CEPRMarkus K Brunnermeier, Princeton University and CEPRMichael C Burda, Humboldt Universitaet zu Berlin and CEPRAline Bütikofer, Norwegian School of EconomicsLuis Cabral, New York University and CEPRPaola Conconi, ECARES, Universite Libre de Bruxelles and CEPRGiancarlo Corsetti, University of Cambridge and CEPRFiorella De Fiore, Bank for International Settlements and CEPRMathias Dewatripont, ECARES, Universite Libre de Bruxelles and CEPRJonathan Dingel, University of Chicago Booth School and CEPRBarry Eichengreen, University of California, Berkeley and CEPRSimon J Evenett, University of St Gallen and CEPRMaryam Farboodi, MIT and CEPRAntonio Fatás, INSEAD Singapore and CEPRFrancesco Giavazzi, Bocconi University and CEPRChristian Gollier, Toulouse School of Economics and CEPRRachel Griffith, IFS, University of Manchester and CEPRTimothy J. Hatton, University of Essex and CEPR
Ethan Ilzetzki, London School of Economics and CEPRBeata Javorcik, EBRD and CEPRSebnem Kalemli-Ozcan, University of Maryland and CEPR Rik FrehenErik Lindqvist, Swedish Institute for Social Research (SOFI)Tom Kompas, University of Melbourne and CEBRAMiklós Koren, Central European University and CEPRAnton Korinek, University of Virginia and CEPRPhilippe Martin, Sciences Po and CEPRWarwick McKibbin, ANU College of Asia and the PacificKevin Hjortshøj O’Rourke, NYU Abu Dhabi and CEPREvi Pappa, European University Institute and CEPRBarbara Petrongolo, Queen Mary University, London, LSE and CEPRRichard Portes, London Business School and CEPRCarol Propper, Imperial College London and CEPRLucrezia Reichlin, London Business School and CEPRRicardo Reis, London School of Economics and CEPRHélène Rey, London Business School and CEPRDominic Rohner, University of Lausanne and CEPRPaola Sapienza, Northwestern University and CEPRMoritz Schularick, University of Bonn and CEPRPaul Seabright, Toulouse School of Economics and CEPRFlavio Toxvaerd, University of CambridgeChristoph Trebesch, Christian-Albrechts-Universitaet zu Kiel and CEPRKaren-Helene Ulltveit-Moe, University of Oslo and CEPRJan C. van Ours, Erasmus University Rotterdam and CEPRThierry Verdier, Paris School of Economics and CEPR
EthicsCovid Economics will feature high quality analyses of economic aspects of the health crisis. However, the pandemic also raises a number of complex ethical issues. Economists tend to think about trade-offs, in this case lives vs. costs, patient selection at a time of scarcity, and more. In the spirit of academic freedom, neither the Editors of Covid Economics nor CEPR take a stand on these issues and therefore do not bear any responsibility for views expressed in the articles.
Submission to professional journalsThe following journals have indicated that they will accept submissions of papers featured in Covid Economics because they are working papers. Most expect revised versions. This list will be updated regularly.
American Economic Review American Economic Review, Applied EconomicsAmerican Economic Review, InsightsAmerican Economic Review, Economic Policy American Economic Review, Macroeconomics American Economic Review, Microeconomics American Journal of Health EconomicsCanadian Journal of EconomicsEconometrica*Economic JournalEconomics of Disasters and Climate ChangeInternational Economic ReviewJournal of Development Economics
Journal of Econometrics*Journal of Economic GrowthJournal of Economic TheoryJournal of the European Economic Association*Journal of FinanceJournal of Financial EconomicsJournal of International EconomicsJournal of Labor Economics*Journal of Monetary EconomicsJournal of Public EconomicsJournal of Political EconomyJournal of Population EconomicsQuarterly Journal of Economics*Review of Economics and StatisticsReview of Economic Studies*Review of Financial Studies
(*) Must be a significantly revised and extended version of the paper featured in Covid Economics.
Covid Economics Vetted and Real-Time Papers
Issue 38, 16 July 2020
Contents
Mutual fund performance and flows during the COVID-19 crisis 1Luboš Pástor and M. Blair Vorsatz
COVID-19 in emerging markets: Firm-survey evidence 37Thorsten Beck, Burton Flynn and Mikael Homanen
Financial markets and news about the coronavirus 68Harry Mamaysky
COVID-19, race, and redlining 129Graziella Bertocchi and Arcangelo Dimico
Unequal consequences of Covid-19 across age and income: Representative evidence from six countries 196Michèle Belot, Syngjoo Choi, Egon Tripodi, Eline van den Broek-Altenburg, Julian C. Jamison and Nicholas W. Papageorge
Consumption and geographic mobility in pandemic times: Evidence from Mexico 218Raymundo M. Campos-Vazquez and Gerardo Esquivel
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Covid Economics Issue 38, 16 July 2020
Copyright: Luboš Pástor and M. Blair Vorsatz
Mutual fund performance and flows during the COVID-19 crisis1
Luboš Pástor2 and M. Blair Vorsatz3
Date submitted: 10 July 2020; Date accepted: 10 July 2020
We present a comprehensive analysis of the performance and flows of U.S. actively managed equity mutual funds during the COVID-19 crisis of 2020. We find that most active funds underperform passive benchmarks during the crisis, contradicting a popular hypothesis. Funds with high sustainability ratings perform well, as do funds with high star ratings. Fund outflows largely extend pre-crisis trends. Investors favor funds that apply exclusion criteria and funds with high sustainability ratings, especially environmental ones. Our finding that investors remain focused on sustainability during this major crisis suggests they view sustainability as a necessity rather than a luxury good.
1 We are grateful to the University of Chicago Booth School of Business for research support. The views in this paper are the responsibility of the authors, not the institutions they are aliated with.
2 Charles P. McQuaid Professor of Finance, University of Chicago Booth School of Business. CEPR Research Fellow. NBER Research Associate and Board member, National Bank of Slovakia.
3 Ph.D. student in Finance, University of Chicago Booth School of Business.
Cov
id E
cono
mic
s 38,
16
July
202
0: 1
-36
1
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
1. Introduction
Active equity mutual funds are well known to have underperformed passive benchmarks, net
of fees.1 Despite its long-lasting underperformance, the active management industry remains
large, managing tens of trillions of dollars. The existence of a large underperforming industry
appears puzzling because an alternative—passive funds—is easily available to investors.
One popular hypothesis is that investors are willing to tolerate this underperformance
because active funds outperform in periods that are particularly important to investors. This
hypothesis is first formulated by Moskowitz (2000) who asks whether mutual funds provide
a hedge against recessions. Glode (2011) formalizes this hypothesis by building a model in
which a fund manager generates active returns that depend on the state of the economy. In
equilibrium, the manager chooses to work harder in periods when investors’ marginal utility
of consumption is higher because investors are willing to pay for this insurance. If active
funds deliver high returns in periods when investors need them the most then these funds’
unconditional performance understates the funds’ true abilities.
We test this hypothesis by analyzing the performance of active mutual funds during the
COVID-19 crisis of 2020. This crisis is particularly suitable for the task at hand for two
reasons. First, it has led to an unprecedented output contraction and the fastest increase in
unemployment on record. Investors surely want to hedge against such a severe crisis. Second,
active managers have an opportunity to perform well during this crisis because the crisis has
created unusually large price dislocations in financial markets. In the equity market, the S&P
500 index experienced its steepest descent in living memory, losing 34% of its value in the five-
week period between February 19 and March 23, 2020 before bouncing back by over 30% by
the end of April. The sharp response of equity markets to COVID-19 is analyzed in a growing
number of studies.2 In the bond market, liquidity evaporated in March 2020, not only for
corporate bonds (e.g., Kargar et al., 2020, and O’Hara and Zhou, 2020) but also for the
usually-liquid Treasuries (e.g., Schrimpf, Shin, and Sushko, 2020). Until liquidity improved
following the interventions from the Federal Reserve, its temporary shortage created massive
market disruptions. For example, in the corporate bond market, bonds traded at large
discounts to credit default swaps, and ETFs traded at large discounts to net asset values
(Haddad, Moreira, and Muir, 2020). In addition, the Treasury market witnessed significant
1See Jensen (1968), Elton, Gruber, Das, and Hlavka (1993), Malkiel (1995), Gruber (1996), Carhart(1997), Wermers (2000), Pastor and Stambaugh (2002), and Fama and French (2010), and others.
2For evidence at the aggregate stock market level, see, for example, Alfaro et al. (2020) and Gormsenand Koijen (2020). For cross-sectional evidence, see Bretscher, Hsu, and Tamoni (2020), Ding et al. (2020),Fahlenbrach, Rageth, and Stulz (2020), Gerding, Martin, and Nagler (2020), Pagano, Wagner, and Zechner(2020), Ramelli and Wagner (2020), and others.
2C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
mispricing between bonds and bond futures (Schrimpf, Shin, and Sushko, 2020). These price
dislocations are due to a combination of factors including record-high volatility and traders
working from home. Under the hypothesis that active funds outperform during recessions,
they should find it particularly easy to outperform when markets are rife with mispricing.
Contrary to this hypothesis, we find that active funds underperform their passive bench-
marks during the COVID-19 crisis. We define the crisis period as the ten-week period
between February 20 and April 30, 2020. We choose February 20 as the starting date be-
cause the stock market peaked on February 19 before its rapid descent. We choose April
30 as the ending date because it is a month-end by which the market largely rebounded,
and also because it puts the market bottom on March 23 roughly in the middle of the crisis
period. The ten-week crisis period is thus roughly evenly split between the crash and the
recovery. Our evidence is based on daily returns of all U.S. active equity mutual funds.
The underperformance of active funds is particularly strong when measured relative to
the S&P 500 benchmark. We find that 74.2% of active funds—about three quarters!—
underperform the S&P 500 during the COVID-19 crisis. The average fund underperformance
is −5.6% (t = −5.37) during the ten-week period, or −29.1% on an annualized basis.
While the S&P 500 is the most popular benchmark among U.S. equity funds, it is not
appropriate for all funds given its large-cap focus. We consider three types of benchmarks
that are tailored to each fund’s investment style: Morningstar-designated FTSE/Russell
benchmarks, fund-designated prospectus benchmarks, and factor-model benchmarks. We
find that active funds also underperform these fund-specific benchmarks, although by lower
margins. For example, 57.6% of funds underperform their FTSE/Russell benchmarks and
54.2% of funds underperform their prospectus benchmarks. The average fund underper-
formance relative to the FTSE/Russell benchmark is −2.1% (t = −3.90) during the crisis
period, or −11% on an annualized basis. Relative to the prospectus benchmark, the average
underperformance is −1.5% (t = −2.49) during the crisis, or −7.7% annualized.
Besides benchmark-adjusted fund returns, we also examine factor-adjusted returns by
computing fund alphas relative to five different factor models. All five alphas are significantly
negative on average during the crisis period, ranging from −7.6% annualized (t = −3.25) for
the six-factor model that includes the five factors of Fama and French (2015) plus momentum,
to −29.1% annualized (t = −7.02) for the CAPM. The fraction of funds with negative alphas
ranges from 60.4% for the four-factor Carhart (1997) model to a stunning 80.2% for the
CAPM. In short, active funds perform poorly during the COVID-19 crisis.
Prior tests of the same hypothesis arrive at a different conclusion. Moskowitz (2000)
3C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
shows that active funds’ returns from 1975 to 1994 are higher during recessions by 6% per
year, on average. Kosowski (2011) analyzes the period from 1962 to 2005 and finds that
mutual fund alphas in recessions exceed those in expansions by 3% to 5% per year, on
average. Glode (2011) reports that funds with poor unconditional performance generate
countercyclical risk-adjusted returns in 1980 through 2005. Kacperczyk, van Nieuwerburgh,
and Veldkamp (2016) find that fund alphas are 1.6% to 4.6% per year higher in recessions
over the 1980–2005 period. Unlike our study, all of these studies examine periods in which
recessions are substantially milder than the COVID-19 crisis.
While active funds as a whole underperform, their performance during the COVID-19
crisis exhibits substantial heterogeneity. One of the strongest predictors of performance is
the sustainability rating from Morningstar. Morningstar assigns between one and five sus-
tainability “globes” to each fund, with more globes denoting higher sustainability. We find
that funds with more globes as of January 31, 2020 have higher benchmark-adjusted returns
between February 20 and April 30, 2020. Remarkably, the relation is monotonic across the
globe categories: five-globe funds outperform four-globe funds, which in turn outperform
three-globe funds, etc. High-globe funds (those with four or five globes) significantly out-
perform the remaining funds within the same investment style by 14.2% per year (t = 4.85)
in terms of FTSE/Russell benchmark-adjusted returns. This result is driven largely by en-
vironmental sustainability—funds with higher environmental ratings outperform those with
lower ratings.
Our findings linking fund performance to sustainability resemble those of Nofsinger and
Varma (2014) who find that socially responsible mutual funds tend to outperform during
periods of market crises. Their findings are based on a sample of 240 U.S. domestic equity
mutual funds in the period of 2000 through 2011, which includes two recessions (2001 and
2007–2009). We examine only one recession (2020) but many more funds. Another related
study, Albuquerque et al. (2020), finds that U.S. firms with high environmental and so-
cial ratings earn comparatively high stock returns in the first quarter of 2020. Ding et al.
(2020) report a similar finding based on corporate social responsibility ratings of firms in 56
countries.3 Our fund-level evidence complements their stock-level evidence in highlighting
the role of sustainability during the COVID-19 crisis. The high returns of sustainable funds
and stocks suggest that market participants’ tastes continue to shift toward green assets and
green products during this crisis (Pastor, Stambaugh, and Taylor, 2020).
Besides sustainability ratings, another strong predictor of fund performance during the
3These findings echo those based on the 2008–2009 recession. Lins, Servaes, and Tamayo (2017) showthat U.S. firms with higher environmental and social ratings perform better during that recession.
4C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
COVID-19 crisis is the fund’s star rating from Morningstar. Star ratings assigned as of
January 31, 2020 predict performance between February 20 and April 30, 2020 positively
and significantly. Similar to Morningstar globes, the relation is monotonic: five-star funds
outperform four-star funds, which outperform three-star funds, etc. Five-star funds signifi-
cantly outperform one-star funds in terms of cumulative benchmark-adjusted returns. One
additional star is associated with an increase in performance of 5.78% per year (t = 2.84) in
terms of FTSE/Russell benchmark-adjusted returns. That is, a five-star fund outperforms
a one-star fund of the same style by about 23% per year, on average.
Finally, we find that growth funds outperform value funds. This finding is only partly
driven by the well-known fact that the growth style outperforms the value style during the
crisis because we measure fund performance net of the fund’s style. In other words, we find
that growth funds beat value funds on a style-adjusted basis. This result is strong when
the style adjustment is performed through a factor model but it is insignificant when the
adjustment is based on the style benchmark. The mixed nature of this evidence suggests roles
for both active management and the superior performance of the growth style in explaining
the different performance of growth and value funds during the crisis.
In addition to fund performance, we analyze capital flows in and out of active mutual
funds. During the COVID-19 crisis, active funds experience steady outflows that largely
continue long-term trends. The outflows are rapid during the market crash but they continue,
albeit at a slower pace, during the market rebound after March 23, 2020.
Fund flows vary substantially across funds. Similar to performance, crisis-period flows
are predictable by funds’ pre-crisis sustainability ratings. Flows are near-monotonic across
the five globe categories, with five-globe funds having the largest net flows and one-globe
funds having the lowest flows between February 20 and April 30, 2020. In particular, one-
globe funds suffer outflows of 2.6% of assets under management over the ten-week period,
whereas five-globe funds’ net flows are roughly zero. This difference, which is statistically
significant, is driven especially by environmental concerns. Furthermore, funds that apply
exclusion criteria in their investment process receive net inflows during the crisis, whereas
funds that do not apply exclusions experience outflows. It is well known that mutual fund
investors have come to favor sustainability-oriented funds in the 2010s (e.g., Bialkowski and
Starks, 2016, and Hartzmark and Sussman, 2019). We find that this pre-crisis trend toward
sustainability continues during the COVID-19 crisis.
A popular perspective in traditional neoclassical economics is that sustainability issues,
such as environmental quality, are “luxury goods” that are likely to be of concern only
5C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
to those whose more basic needs for food, housing, and survival are adequately met (e.g.,
Baumol and Oates, 1979).4 This perspective predicts that interest in sustainability should
subside during a major economic and health crisis. In contrast, we find that investors retain
their commitment to sustainability during the COVID-19 crisis. This finding suggests that
investors have come to view sustainability as a necessity rather than a luxury good.
The performance hypothesis rejected by our evidence—that active funds outperform in
recessions—is not the only possible explanation for why active management remains popu-
lar despite its poor track record. Gruber (1996) suggests that some investors suboptimally
rely on active management because they are influenced by advertising, brokers, institutional
arrangements, or tax considerations. Pastor and Stambaugh (2012) argue that a large active
management industry can be rationalized if investors believe that active managers face de-
creasing returns to scale. In their model, rational investors respond to past underperformance
of active funds by withdrawing money, which improves those funds’ future performance to
the point where investors are indifferent between investing actively or passively.
Our focus on crisis-period fund performance is also related to the literature on time-
varying fund manager skill. An important early contribution is Ferson and Schadt (1996).
More recently, Kacperczyk, van Nieuwerburgh, and Veldkamp (2016) develop a model of
optimal attention allocation over the business cycle. In their model, fund managers allocate
more attention to idiosyncratic shocks in expansions and to aggregate shocks in recessions.
Similarly, Kacperczyk, van Nieuwerburgh, and Veldkamp (2014) find that fund managers
exhibit better stock picking in expansions and better market timing in recessions. We do not
attempt to separate stock selection from market timing during the COVID-19 crisis because
such an exercise would require time series of fund holdings, which are widely available only
on a quarterly basis. Data availability also limits our ability to test the hypothesis that the
profit opportunities created by COVID-19 lead active funds to trade more, improving their
future performance (Pastor, Stambaugh, and Taylor, 2017). Whether funds increase their
turnover in 2020, and whether this turnover causes better future fund performance, remains
to be seen because the turnover data from the SEC are only annual.
The paper is organized as follows. In Section 2, we describe our data. In Section 3, we
analyze fund performance and its determinants. In Section 4, we discuss fund flows and their
determinants. Section 5 concludes. Additional empirical results and database construction
details are located in the Appendix, which is available on the authors’ websites.
4An example of this common view is the controversial “Summers memo” from 1991, in which the WorldBank’s Chief Economist suggests that the Bank should be encouraging more migration of dirty industriesto the least-developed countries. One of the reasons given in the memo is that “the demand for a cleanenvironment for aesthetic and health reasons is likely to have very high income elasticity.”
6C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
2. Data
We use daily data from Morningstar Direct covering the period from January 1, 2017 to
April 30, 2020. Our original sample covers 4,292 actively managed U.S. equity mutual
funds, although we primarily focus on the 3,626 funds with at least one non-missing net
return between February 20 and April 30, 2020. The latter sample represents $4.9 trillion of
total net assets as of January 31, 2020.
Our fund universe is constructed largely following Pastor, Stambaugh, and Taylor (2015),
with two main differences. First, we also include international and sector equity funds domi-
ciled in the U.S. Second, we do not require funds to appear in both CRSP and Morningstar;
we use only Morningstar data. As in Pastor, Stambaugh, and Taylor (2015), we use the
Morningstar FundID variable to aggregate share classes to the fund level.5
We use keywords in the Morningstar Category variable and the prospectus benchmark
to exclude bond funds, money market funds, real estate funds, target retirement funds, and
other non-equity funds. We also exclude funds identified by Morningstar as passive index
funds and funds whose name contains the word “index.” In our baseline results, we also use
a fund size filter to include only funds with at least $15 million of net assets on January
31, 2020. Excluding the smallest funds has been advocated by Elton, Gruber, and Blake
(2001), among many others. This screen is particularly relevant for fund flows because
modest dollar flows can translate into extreme percentage flows for the smallest funds. This
subsample covers 2,764 funds and $4.891 trillion of total net assets.
Throughout our analysis, we use funds’ returns net of the expense ratio because our
goal is to measure the return delivered to clients after fees. Despite being very fresh (we
downloaded data through April 30, 2020 in May 2020), the data appear to be free of salient
errors. For example, none of our sample’s 2,692,799 fund-level daily net return observations
are below -90% and only one is larger than 100%. We adjust fund returns for daily benchmark
returns, also obtained from Morningstar, and for daily factor returns, which we obtain from
Ken French’s data library along with the risk-free rate.
We rely on three main Morningstar categorization variables throughout our analysis: the
5Many funds have multiple share classes, which are tied to the same pool of assets but have differentfee structures. Since different share classes of the same fund have the same Morningstar FundID value, wecan use the FundID variable to aggregate the share classes up to the fund level. Specifically, we computea fund’s total net assets by summing total net assets across the fund’s share classes, setting the fund-levelvariable to missing if total net assets are missing for any of the share classes on that date. The fund’s netreturns, net expense ratio, and turnover ratio are averaged (lag-asset-weighted) across all share classes withnon-missing values.
7C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Morningstar Category, the Morningstar Institutional Category, and the Global Category. In
our full sample of 3,626 funds, there are 39 Global Categories, 52 Morningstar Categories, and
93 Morningstar Institutional Categories. Each of these variables classifies a fund based on its
investment style, sector, and geographical orientation. The Global Categories are the coarsest
classification system and are used by Morningstar as groupings within which sustainability
can be ranked. These categories include U.S. equity large-cap blend, U.S. equity small-cap,
energy sector equity, and Latin America equity. We use the Global Categories for style fixed
effects. The Morningstar Category variable is built on the 3-by-3 style box of size tilts (large-
cap vs. small-cap) and growth vs. value style tilts. We use this variable to follow Pastor,
Stambaugh, and Taylor (2015) in their classification of funds into equity and non-equity
categories. Morningstar uses these groupings to rank performance in terms of star ratings.
Last, the Morningstar Institutional Category variable is built on an extended version of the
3-by-3 style box with size tilts including micro-cap and giant and style tilts including deep
value and high growth. We use this finest classification system for clustering our standard
errors. While this is conservative relative to the more standard treatment of clustering at
the fund level, we believe this appropriately accounts for how the health crisis shock may
generate residual correlation among funds with similar strategies. For further details of our
data construction, see the Appendix.
3. Fund Performance
Figure 1 provides a preliminary look at the performance of active funds during the COVID-
19 crisis, along with the performance of the most popular passive benchmark: the S&P 500
index. We normalize the levels of both the S&P 500 and each fund’s net asset value to 100
as of February 19, 2020. For each day t after February 19, we compute the price indices for
each fund as well as the S&P 500 by compounding the corresponding daily returns:
Ft = 100(1 + rF1 )(1 + rF2 )...(1 + rFt ) (1)
Bt = 100(1 + rB1 )(1 + rB2 )...(1 + rBt ) , (2)
where Ft is the fund price index, Bt is the price index for the passive benchmark, rFt is the
fund’s net return on day t, and rBt is the benchmark return. Figure 1 plots both Bt and the
average value of Ft across all funds. The figure also plots a 95% confidence interval around
average Ft. Standard errors are clustered on the Morningstar Institutional Category, both
here and in all subsequent figures reporting confidence intervals.
8C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 1. Average Fund Performance vs. the S&P 500 During the Crisis. This figure plots theperformance of the average active equity mutual fund against the S&P 500 in February 20 through April30, 2020. Both price indices are initialized at 100 on February 19, 2020 and computed by compoundingdaily returns. The fund average is computed by adding the average difference between the fund price indexand the S&P 500 price index to the S&P 500 price index. Standard errors are estimated for this differ-ence and are clustered on the Morningstar Institutional Category. 95% confidence intervals are plotted in red.
Computing Ft in equation (1) requires all of the fund’s daily returns starting February
20 through day t. Any gap in the fund’s return series, however short, would necessitate the
fund’s deletion from the average calculation. To avoid deleting too many funds, we replace
any missing returns by the average return across all funds with the same FTSE/Russell
9C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
benchmark on the same day, thus preserving the average level of performance across funds.
We only replace missing returns for which there exists a non-missing return later in the
fund’s history by April 30. That is, we do not replace any missing returns at the end of a
fund’s history because funds that stop reporting returns to Morningstar may no longer be
alive. Altogether, we replace 19,124 missing returns, which account for 13.9% of our sample.
We apply this “patch” not only in Figure 1 but also in Figures 2 through 8.6 We do not
replace missing returns in Tables 1 through 3 because the analysis behind those tables does
not require funds to have continuous return series.
Figure 1 shows that the S&P 500 loses 34% of its value between February 19 and March
23, before gaining 30% by April 30, 2020. The average active fund performs similarly, but it
significantly underperforms the S&P 500 during the crisis. The April 30 price index levels
are 86.01 for the S&P 500 but only 82.45 for the average active fund.
Given its focus on large-cap stocks, the S&P 500 is not the most appropriate benchmark
for every fund. Several large-cap technology stocks performed well during the crisis, making
the S&P 500 hard to beat. We thus compare each fund’s return also to the returns of
two benchmarks tailored to the fund’s investment style: the prospectus benchmark and
the FTSE/Russell benchmark. The prospectus benchmark is chosen by the fund itself (with
some potential for strategic choice, as discussed by Sensoy, 2009), whereas the FTSE/Russell
benchmark is assigned to each fund by Morningstar based on the fund’s holdings.
Figure 2 compares fund performance to the FTSE/Russell benchmark (Panel A), the
prospectus benchmark (Panel B), and the S&P 500 (Panel C). Unlike Figure 1, which plots
index levels, Figure 2 plots the cumulative performance of the average active fund relative
to the benchmark. Specifically, at each date t after February 19, 2020, the figure plots the
average value of log(Ft) − log(Bt), where Ft and Bt are defined in equations (1) and (2).
Figure 2 shows that active funds significantly underperform their benchmarks, on average.
As of April 30, 2020, the average underperformance over the ten-week period is 1.53% relative
to the FTSE/Russell benchmark, 0.94% relative to the prospectus benchmark, and 4.77%
relative to the S&P 500. This underperformance is highly statistically significant in Panel C,
and marginally significant in Panels A and B. Moreover, Figure 2 underestimates the actual
underperformance due to a mild survivorship bias because average Ft can only be computed
across funds that have survived through time t. During the ten-week period, 22 funds drop
out of our sample, so their returns are excluded from the plot as of April 30.
6All figures look virtually identical if we replace missing returns not by average fund returns but ratherby the returns on the fund’s FTSE/Russell benchmark on the same day.
10C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 2. Average Benchmark-Adjusted Fund Performance. This figure plots the cumulativecompound performance of the average active equity mutual fund in February 20 through April 30, 2020relative to three benchmarks: the Morningstar-designated FTSE/Russell benchmark (Panel A), theprospectus benchmark (Panel B), and the S&P 500 (Panel C). Relative performance is measured bylog(Ft) − log(Bt), where Ft and Bt are the cumulative compounded daily returns of the average fund andthe benchmark, respectively. Standard errors are estimated for this difference and are clustered on theMorningstar Institutional Category. 95% confidence intervals are plotted in red.
Table 1 reports average benchmark-adjusted fund performance in a way that is immune
to the survivorship bias. For each fund, live or dead, we take all of the fund’s available daily
returns in the given time period and subtract the same days’ returns on the corresponding
benchmark. We report annualized averages of those benchmark-adjusted returns in the first
three columns of Panel A of Table 1. The average fund underperforms its FTSE/Russell
11C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
benchmark by 11.02% per year, with a t-statistic of −3.90. Average underperformance rel-
ative to the prospectus benchmark is slightly smaller, 7.70% per year, but still significant
(t = −2.49). The average fund underperforms the S&P 500 by a whopping 29.12% per year
(t = −5.37). We observe underperformance not only during the full ten-week crisis period
but also during both subperiods, the first of which captures the market crash and the second
the recovery, and also in the pre-crisis period (October 1, 2019 to January 31, 2020).
Table 1Fund Performance
This table describes active equity mutual funds’ performance against both benchmarks and factor models.Panel A reports simple averages across funds of estimated deltas and alphas, all reported in annualizedpercentage terms. The deltas are average differences between the fund’s net returns and its benchmarkreturns. The benchmarks are the FTSE/Russell benchmark, the prospectus benchmark, and the S&P 500.The alphas are estimated intercepts from the regressions of excess net fund returns on factor returns. Thefactor models are described in the text. Panel B reports the fraction of funds that underperform (i.e., havea negative delta or alpha). The time periods are: crisis (February 20 to April 30, 2020); crash (February 20to March 23, 2020); recovery (March 24 to April 30, 2020); and pre-crisis (October 1, 2019 to January 31,2020). Standard errors are clustered on the Morningstar Institutional Category. t-statistics are in brackets.
(1) (2) (3) (4) (5) (6) (7) (8)
∆FTSE/RusBench ∆Prosp
Bench ∆S&P500Bench αCAPM αFF3 αCar4 αFF5 αFF5+Mom
Panel A. Average Fund Performance (%)
Crisis -11.02 -7.70 -29.12 -29.11 -11.30 -7.84 -9.90 -7.62[-3.90] [-2.49] [-5.37] [-7.02] [-5.16] [-3.22] [-4.51] [-3.25]
Crash -7.91 -8.19 -64.31 -80.94 -37.75 -38.98 -49.84 -51.11[-1.54] [-1.33] [-4.77] [-8.33] [-4.25] [-3.90] [-5.91] [-5.77]
Recovery -12.68 -7.55 -5.81 8.47 17.49 20.76 18.75 22.83[-3.94] [-2.13] [-1.14] [1.46] [2.83] [4.12] [3.00] [4.42]
Pre-Crisis -2.11 -1.28 -3.03 -5.14 -3.12 -3.01 -2.65 -2.59[-2.86] [-2.09] [-1.80] [-3.87] [-4.37] [-3.64] [-3.22] [-2.88]
Panel B. Fraction of Funds Underperforming (%)
Crisis 57.59 54.17 74.24 80.15 69.66 60.35 67.80 60.43Crash 51.62 48.57 63.48 83.79 78.57 80.01 80.65 81.41
Recovery 55.73 55.64 55.77 53.09 39.53 34.79 40.08 34.28Pre-Crisis 63.58 59.85 67.26 71.20 73.31 71.54 70.02 69.51
The remaining columns of Panel A of Table 1 report average fund alphas from five
multifactor models: the capital asset pricing model (CAPM), the three-factor model of Fama
and French (1993), the four-factor model of Carhart (1997), the five-factor model of Fama
12C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
and French (2015), and a six-factor model that includes those five factors plus momentum.
For a fund’s alpha to be included in the average, the fund must have at least 15 non-missing
net returns for the time period of interest. All five alphas are significantly negative during the
crisis period, ranging from −7.62% per year (t = −3.25) for the six-factor model to −29.11%
per year (t = −7.02) for the CAPM. The alphas are particularly negative during the crash
period (February 20 to March 23, 2020), ranging from −37.75% to −80.94% per year across
the five models.
Panel B of Table 1 shows that 57.6% of funds underperform their FTSE/Russell bench-
marks during the crisis. Additionally, 54.2% of funds underperform prospectus benchmarks
and 74.2% of funds—almost three quarters!—underperform the S&P 500. More than 80%
of funds have negative CAPM alphas during the crisis period. The fraction of funds with
negative alphas ranges from 60.4% to 80.2% across the five models. Regardless of how we
look at the data, we see active funds underperforming during the crisis.
3.1. Sustainability
We find that funds with higher sustainability ratings perform better during the crisis. For
each fund, Morningstar evaluates how well the fund’s holdings perform on ESG issues relative
to the fund’s peer group (i.e., Morningstar Global Category). Morningstar uses company-
level ESG scores from Sustainalytics to determine each fund’s asset-weighted average un-
managed ESG risk exposure. Then, within each peer group, these scores are fitted to an
approximate normal distribution to award 1, 2, 3, 4, or 5 sustainability globes to each fund.7
Funds with 5 globes are the most sustainable and funds with 1 globe are the least sustainable.
We find that funds with more globes perform better during the crisis.
Figure 3 shows the distributions of cumulative returns during the crisis across funds with
different sustainability ratings, which are assigned as of January 2020. We collect funds in
two groups: funds with 4 or 5 globes (“high sustainability”) and funds with 1 or 2 globes
(“low sustainability”). Panel A shows the distributions of cumulative total fund returns,
whereas Panel B shows cumulative benchmark-adjusted returns, which are adjusted using
FTSE/Russell benchmarks. Specifically, Panel A shows log(Ft), where Ft is in equation (1)
and t corresponds to April 30, 2020, while Panel B shows log(Ft) − log(Bt), where Bt is in
equation (2) for the fund’s FTSE/Russell benchmark. Both panels clearly show that more
sustainable funds perform better in the crisis.
7Within each peer group, the top 10% of funds receive 5 globes, the next 22.5% receive 4 globes, the next35% receive 3 globes, the next 22.5% receive 2 globes, and the bottom 10% receive 1 globe.
13C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 3. Cumulative Return Densities Across Sustainability Ratings. This figure plots densitiesof funds’ cumulative returns from February 20 to April 30, 2020 for two categories of sustainability: high(four or five Morningstar globes) and low (one or two globes), both assigned as of January 2020. In PanelA, the cumulative returns are unadjusted, given by log(Ft) where Ft = (1 + r
F1 )(1 + r
F2 ) . . . (1 + r
Ft ) is
the fund’s cumulative total return. In Panel B, the cumulative returns are benchmark-adjusted, given bylog(Ft) − log(Bt), where Bt is the cumulative total return of the fund’s FTSE/Russell benchmark.
Figure 4 presents the sustainability result from a perspective similar to Figure 2, plotting
cumulative fund performance relative to the benchmark, or log(Ft) − log(Bt). We consider
the same three benchmarks as before: FTSE/Russell (Panels A and B), prospectus (Panels
C and D), and the S&P 500 (Panels E and F). In the left panels (A, C, and E), we plot the
average cumulative performance differences for each of the five globe groups. In the right
panels (B, D, and F), we report 95% confidence intervals for high-sustainability funds (4 or
5 globes) and low-sustainability funds (1 or 2 globes).
14C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 4. Benchmark-Adjusted Fund Performance: Sustainability Ratings. This figure plotsthe cumulative compound performance in February 20 through April 30, 2020 for fund categories withdifferent numbers of Morningstar sustainability globes assigned as of January 2020. “High sustainability”denotes funds with 4 or 5 globes while “low sustainability” denotes funds with 1 or 2 globes. Performance ismeasured relative to the FTSE/Russell benchmark (Panels A and B), the prospectus benchmark (Panels Cand D), and the S&P 500 (Panels E and F). Relative performance is measured by log(Ft)− log(Bt), where Ft
and Bt are the cumulative compounded daily returns of the average fund and the benchmark, respectively.Standard errors are estimated for this difference and are clustered on the Morningstar Institutional Category.
Remarkably, Figure 4 shows a monotonic relation between benchmark-adjusted fund per-
formance and sustainability globes: five-globe funds outperform four-globe funds, which out-
perform three-globe funds, which in turn outperform two-globe funds, which beat one-globe
funds. This monotonicity is present for all three benchmarks. The performance difference
between high-sustainability funds and low-sustainability funds is marginally statistically sig-
nificant.8 The significance is stronger in the subsequent regression analysis in Table 2.
8This is a difference-in-difference type of calculation, where one difference is between the fund and its
15C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 5. Benchmark-Adjusted Fund Performance: Sustainability Components. This figureplots the cumulative compound performance in February 20 through April 30, 2020 for fund categorieswith different Morningstar sustainability scores. These scores represent historical portfolio sustainabilityscores (Panel A), environmental scores (Panel B), social scores (Panel C), and governance scores (Panel D).For each of the four scores, the top 30% most sustainable funds are labeled as “greener” and the bottom30% of funds are labeled “browner.” Performance is measured relative to the FTSE/Russell benchmark bylog(Ft) − log(Bt), where Ft and Bt are the cumulative compounded daily returns of the average fund andthe benchmark, respectively.
Given the important role of sustainability in determining fund performance, we inves-
tigate which dimensions of sustainability—E, S, or G—matter the most during the crisis.
After sorting funds by their individual E, S, and G scores from Morningstar, we separate
funds into two groups, “greener” (top 30%) and “browner” (bottom 30%), for each of the
three scores. We do the same for the composite historical sustainability score, based on
benchmark and the other difference is between funds with different numbers of globes.
16C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
which Morningstar assigns globes to each fund.9 We perform the greener-versus-browner
comparisons in Figure 5, whose four panels are analogous to Panel B of Figure 4, except
that sustainability globes are replaced by the four metrics described above. In all four panels,
we benchmark funds against FTSE/Russell.
Figure 5 shows that funds with high sustainability scores outperform those with low
scores. This result from Panel A is not surprising, given the prior results from Figure 4.
More interesting, funds with high environmental (E) scores outperform those with low E
scores (Panel B), whereas funds with high social (S) scores underperform those with low
S scores (Panel C). According to Panel D, funds’ governance (G) scores have no effect on
performance. To make the figure easy to read, we do not show confidence intervals, but we
do show them in the Appendix. Only the pattern in Panel B is statistically significant.
Figures 3 through 5 demonstrate that more sustainable funds perform better during the
crisis. We further examine this result by conducting regression analysis, with two benefits.
First, regressions allow us to see whether the result survives the inclusion of many control
variables. Second, we remove the slight survivorship bias discussed earlier.
Table 2 analyzes the determinants of crisis-period fund performance in cross-sectional
regressions with large numbers of controls. Panel A focuses on benchmark-adjusted perfor-
mance, using FTSE/Russell benchmarks. Panel B considers factor-adjusted performance,
using the four-factor Carhart (1997) model. The right-hand-side variables include indicators
for sustainability, exclusions, and the growth investment style, as well as the Morningstar
star rating. Fund-level controls include the log of fund age, the log of the fund’s total net
assets, fund turnover, expense ratio, cash position, the Morningstar medal rating, and mar-
ket beta (in Panel B only because there is no beta estimation in Panel A). Industry controls
are the fund’s net investment position as a percentage of net assets in industries including
energy, healthcare, and technology, among others. All regressions include style fixed effects,
where style is measured at the level of the Morningstar Global Category. As a result, the
relevant comparisons are across funds within the same investment style.
9A fund’s individual E, S, and G scores do not simply add up to the fund’s historical sustainability score.There does not appear to be a simple direct mapping between the two sets of scores.
17C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 2Determinants of Fund Performance During the Crisis
The table reports slope coefficients estimated from regressions of fund performance in February 20 to April 30,2020 on fund characteristics and controls. In Panel A, the dependent variable is FTSE/Russell-benchmark-adjusted performance; in Panel B, it is the Carhart four-factor alpha. Both performance measures areestimated using simple returns and expressed in annualized percentage terms. Global category fixed effectsare based on the Morningstar Global Category variable. Fund-level controls include the log of the fund’sage in days, the log of the fund’s January 31, 2020 total net assets (TNA), turnover ratio as of January2020, net expense ratio as of January 2020, net cash position (as a percent of TNA) as of January 2020,Morningstar medal rating as of January 2020, and, in Panel B only, market beta estimated from the October1, 2019 to January 31, 2020 period. Industry controls include the fund’s net position as a percent of TNA inbasic materials, communication services, consumer cyclical, consumer defensive, healthcare, industrials, realestate, technology, energy, financial services, and utilities. Standard errors are clustered on the MorningstarInstitutional Category. t-statistics are in brackets.
(1) (2) (3) (4) (5) (6) (7)
Panel A. Benchmark-Adjusted Performance
I(4 or 5 Sustainability Globes) 14.21 11.51 8.61 9.76[4.85] [3.22] [2.26] [2.60]
I(Employs Exclusions) 8.61 5.47 2.03 2.79[3.26] [2.44] [1.05] [1.24]
Star Rating 5.78 5.12 7.00 6.49[2.84] [2.42] [3.50] [3.41]
I(Growth Tilt) 12.43 7.24 9.39 5.15[2.35] [1.16] [1.70] [0.75]
Global Category FE Yes Yes Yes Yes Yes Yes YesFund-Level Controls No No No No No Yes YesIndustry Controls No No No No No No YesObservations 2,494 2,561 2,286 2,561 2,251 1,632 1,604Adjusted R2 0.06 0.05 0.06 0.06 0.06 0.12 0.15
Panel B. Factor-Adjusted Performance
I(4 or 5 Sustainability Globes) 5.59 2.67 3.04 3.47[4.25] [2.39] [2.55] [3.15]
I(Employs Exclusions) −0.89 −2.61 −3.46 −3.16[−0.50] [−1.52] [−2.12] [−2.19]
Star Rating 3.15 2.51 3.25 3.13[3.35] [2.79] [5.92] [5.42]
I(Growth Tilt) 10.62 7.53 7.51 7.77[5.58] [4.09] [3.74] [4.10]
Global Category FE Yes Yes Yes Yes Yes Yes YesFund-Level Controls No No No No No Yes YesIndustry Controls No No No No No No YesObservations 2,233 2,363 2,104 2,363 2,020 1,522 1,494Adjusted R2 0.10 0.12 0.11 0.12 0.10 0.42 0.46
18C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 2 confirms that funds with higher sustainability ratings perform better during the
crisis. As before, we define high-sustainability funds as those with 4 or 5 globes. Column
1 of the table includes no controls other than style fixed effects. In column 1 of Panel
A, the slope on the high-sustainability indicator is 14.21 (t = 4.85), indicating that high-
sustainability funds outperform the remaining funds within the same style by 14.21% per
year during the crisis. The high-sustainability indicator is also highly significant in column
1 of Panel B (t = 4.25), where fund returns are factor-adjusted rather than benchmark-
adjusted. Sustainability thus remains a significant determinant of performance even after
style fixed effects are included. This is not surprising, given our prior results, because both
sustainability ratings and fund returns are style-adjusted, though in slightly different ways—
sustainability ratings by Morningstar, with respect to the Morningstar Global Category, and
returns by us, with respect to the fund’s FTSE/Russell benchmark.
More interesting, sustainability remains significantly associated with fund performance
after the inclusion of fund-level and industry controls. The slope on the high-sustainability
indicator decreases as controls are added, but it remains both statistically and economically
significant even when all controls are included: 9.76 (t = 2.60) in Panel A and 3.47 (t = 3.15)
in Panel B. To summarize, we find that funds with high sustainability ratings perform better
during the crisis.
A subset of funds employ exclusions in their investment process. These funds exclude
from their portfolios stocks of firms such as tobacco producers or gun manufacturers whose
business they deem unacceptable. Exclusions represent one possible approach to sustain-
ability, one that discards the opportunities to engage with the firm as well as to benefit from
the potential mispricing of the firm’s stock. 107 of our funds employ exclusions, representing
3.9% of our TNA-screened sample.
Table 2 shows that funds that employ exclusions outperform same-style funds that do
not employ exclusions by 8.61% per year (t = 3.26) in terms of benchmark-adjusted returns.
However, this result weakens, and eventually loses significance, after adding enough control
variables. The result does not obtain on a factor-adjusted basis; if anything, it goes the
other way (Panel B). The exclusion aspect of sustainability therefore does not have a robust
association with fund performance during the crisis.
19C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
3.2. Star Ratings
Besides sustainability, the most important determinant of active fund performance during the
crisis is the fund’s star rating as of January 31, 2020. To calculate star ratings, Morningstar
computes each fund’s risk-adjusted performance over the prior three, five, and ten years
relative to the fund’s peer group. Averaging across the three periods, Morningstar awards
1, 2, 3, 4, or 5 stars to each fund, with 5 stars going to the best-performing funds.10 We find
that funds with higher star ratings perform better during the crisis.
Figure 6. Cumulative Return Densities Across Star Ratings. This figure plots densities of funds’cumulative returns from February 20 to April 30, 2020 for two categories of star ratings: high (four or fiveMorningstar stars) and low (one or two stars), both assigned as of January 2020. In Panel A, the cumulativereturns are unadjusted, given by log(Ft) where Ft = (1 + r
F1 )(1 + r
F2 ) . . . (1 + r
Ft ) is the fund’s cumulative
total return. In Panel B, the cumulative returns are benchmark-adjusted, given by log(Ft)− log(Bt), whereBt is the cumulative total return of the fund’s FTSE/Russell benchmark.
10As with the sustainability globes, within each peer group, the top 10% of funds receive 5 stars, thenext 22.5% receive 4 stars, the next 35% receive 3 stars, the next 22.5% receive 2 stars, and the bottom10% receive 1 star. A fund must have at least three years of performance to be considered for a rating,and depending on its age, a combination of three-year, five-year, and ten-year performance measures areaveraged to construct the fund’s raw performance score.
20C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 6 shows the distributions of cumulative returns during the crisis across funds with
different star ratings, similar to Figure 3. We collect funds in two groups: funds with 4
or 5 stars (“high”) and funds with 1 or 2 stars (“low”). Panel A shows the distributions
of cumulative total fund returns whereas Panel B shows cumulative returns adjusted for
FTSE/Russell benchmark returns. Both panels clearly show that funds with more stars
perform better during the crisis.
Figure 7. Benchmark-Adjusted Fund Performance: Star Ratings. This figure plots thecumulative compound performance in February 20 through April 30, 2020 for fund categories withdifferent numbers of Morningstar stars assigned as of January 2020. Performance is measured relativeto the FTSE/Russell benchmark (Panels A and B), the prospectus benchmark (Panels C and D), andthe S&P 500 (Panels E and F). Relative performance is measured by log(Ft) − log(Bt), where Ft andBt are the cumulative compounded daily returns of the average fund and the benchmark, respectively.Standard errors are estimated for this difference and are clustered on the Morningstar Institutional Category.
21C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 7 shows the same result from a different perspective. Similar to the layouts from
Figures 2 and 4, Figure 7 plots cumulative benchmark-adjusted fund performance for groups
of funds with different star ratings. The relation between benchmark-adjusted performance
and star ratings is monotonic across the five star groups, with five-star funds performing the
best and one-star funds performing the worst. This striking monotonicity is observed for all
three benchmarks. The figure also shows that five-star funds outperform one-star funds by
a significant margin for all three benchmarks.
Table 2 confirms the important role of the star rating in our regression setting with con-
trols and style fixed effects. The star rating significantly predicts both benchmark-adjusted
and factor-adjusted returns, with t-statistics ranging from 2.42 to 3.50 in Panel A and from
2.79 to 5.92 in Panel B. This is a surprising result—it is not clear a priori why Morningstar
star ratings, which are computed before the crisis from historical risk-adjusted returns, should
have such strong predictive power for fund performance during the crisis. The result is sig-
nificant not only statistically but also economically. For example, the slope coefficient of 5.78
in column 3 of Panel A indicates that one extra star is associated with a higher crisis-period
benchmark-adjusted return of 5.78% per year. Therefore, a five-star fund outperforms a
one-star fund of the same style by four times that amount, 23.1% per year, on average.
3.3. Value versus Growth
Sustainability and stars are the most robust predictors of active fund performance during
the crisis. Next in line, though less robust, is the value/growth investment style. We find
that growth funds tend to outperform value funds. Importantly, we are not saying that the
growth style outperforms the value style during the crisis—that is well known (e.g., HML’s
crisis-period return is −18%). What we are saying is that growth funds deliver higher returns
than value funds on a style-adjusted basis.
To decide which funds follow the value and growth investment styles, we use the equity
style box variable from Morningstar. We define value funds as funds classified as large-cap
value, mid-cap value, or small-cap value. We define growth funds as funds classified as
large-cap growth, mid-cap growth, or small-cap growth.
Figure 8 shows that growth funds outperform value funds, on average, for all three
benchmarks. The outperformance is statistically significant when measured against the S&P
500 and prospectus benchmarks, but not against the FTSE/Russell benchmarks.
22C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 8. Benchmark-Adjusted Fund Performance: Growth vs. Value Funds. This figureplots the cumulative compound performance in February 20 through April 30, 2020 for growth vs valuefunds, as determined by the Morningstar equity style box. Performance is measured relative to theFTSE/Russell benchmark (Panel A), the prospectus benchmark (Panel B), and the S&P 500 (Panel C).Relative performance is measured by log(Ft) − log(Bt), where Ft and Bt are the cumulative compoundeddaily returns of the average fund and the benchmark, respectively. Standard errors are estimated for thisdifference and are clustered on the Morningstar Institutional Category. 95% confidence intervals are shown.
Table 2 finds the same outperformance in our regression setting with controls and style
fixed effects. In Panel B, where fund performance is factor-adjusted, the indicator variable
for the growth tilt is always positive and significant, with t-statistics exceeding 3.7 in all
specifications. For example, using the estimate from column 4, growth funds outperform non-
growth funds by 10.62% per year (t = 5.58) during the crisis. In Panel A, the estimated slopes
on the growth indicator are similar in magnitude to those in Panel B but their statistical
significance is substantially weaker, with t-statistics ranging from 0.75 to 2.35 across the four
specifications.
23C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
3.4. Robustness
Our main results are robust to a variety of methodological modifications. As noted earlier,
our sample is restricted to active equity funds with at least $15 million in total net assets as
of January 31, 2020. However, we show in the Appendix that the results from Table 1 are
extremely similar even if we do not impose this size screen.
Another screen that is commonly imposed on mutual fund samples is an age screen.
Researchers often exclude young funds because of a concern about the incubation bias (Evans,
2010). This bias can appear if researchers analyze historical fund data with a delay that
would allow the bias to creep in. The bias is not a concern in our study because there is
no such delay—we constructed our fund sample in May 2020, shortly after the end of our
sample period. Nonetheless, we show in the Appendix that our main results are extremely
similar also when we exclude funds less than two years old from the sample.
Our tables report evidence based on simple returns. Our plots of cumulative performance
are based on log (i.e., continuously compounded) returns because those returns cumulate over
time in a tractable manner. This distinction is immaterial—our main table results are very
similar if we replace simple returns by log returns, as we show in the Appendix.
To remove the effects of investment style, we include style fixed effects in our regressions.
Nonetheless, we show in the Appendix that the regression results from Table 2 are similar
if style fixed effects are excluded. The Appendix also reports subperiod results for Table 2
and its variations, dividing the full crisis period into the crash period (February 20 to March
23, 2020) and the recovery period (March 24 to April 30, 2020).
Recall that Figure 3 shows the distributions of crisis-period returns across two groups of
funds, those with high sustainability ratings (4 or 5 globes) and low sustainability ratings
(1 or 2 globes). In the Appendix, we present analogous plots showing three distributions
corresponding to funds with 1, 3, and 5 globes, and also five distributions, one for each
possible number of globes. Those plots are more cluttered but they convey the same message
as Figure 3—that more sustainable funds perform better in the crisis.
Similarly, Figure 6 shows the distributions of returns across funds with high star ratings
(4 or 5 stars) and low star ratings (1 or 2 stars). In the Appendix, we present analogous
plots showing three distributions corresponding to funds with 1, 3, and 5 stars, and also
five distributions, one for each possible number of stars. Again, those plots convey the same
message as Figure 6: funds with more stars perform better during the crisis.
24C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
4. Fund Flows
Our key measure for assessing fund flows is the cumulative net fund flow percentage. Daily
net fund flows, in dollars, are computed following Barber, Huang, and Odean (2016) as
FDi,t = TNAi,t − (1 +Ri,t)TNAi,t−1 , (3)
where TNAi,t is the total net assets of fund i on date t and Ri,t is the net return of fund i
on date t. To convert the dollar values of net fund flows into a cumulative percentage, the
values of FDi,t are accumulated across the time period of interest and divided by the total
net assets of fund i on the day before the period of interest begins. Given the sensitivity of
the cumulative net flow percentage to missing values, we restrict consideration to funds with
entirely non-missing daily net fund flows.11
Figure 9. Aggregate Net Fund Flows. This figure plots aggregate net flows into active equity mutualfunds during the crisis period (Panel A) and over the past three years (Panel B). Specifically, Panel A plotstotal cumulative net fund flows (in both USD billions and as a percent of February 19, 2020 aggregate totalnet assets) over the February 20 to April 30, 2020 period. Panel B covers the January 4, 2017 to April 30,2020 period, and it expresses flows as a percent of January 3, 2017 total net assets.
11In our baseline fund size filtered sample, we retain 2,082 funds (75.3%) over the full February 20 to April30, 2020 time period, 2,137 funds (77.3%) over the February 20 to March 23, 2020 time period, and 2,219funds (80.3%) over the March 24 to April 30, 2020 time period.
25C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 9 shows the time series of cumulative net fund flows into active equity mutual
funds, both in dollar terms and in percentage terms. Panel A shows that active funds
experience steady outflows during the COVID-19 crisis of about 43 billion dollars, or 1.3%
of assets under management. The pace of outflows is fairly rapid during the market crash
between February 20 and March 23, 2020. Outflows continue, albeit at a slower pace, after
the market rebound. Panel B shows that between January 2017 and April 2020, active funds
experience outflows amounting to about 5% per year as a fraction of their initial assets. These
steady outflows reflect the well-known ongoing trend toward passive investment management.
Year 2020 does not stand out relative to prior years, indicating that crisis-period outflows
largely extend their long-term pre-crisis trend.
4.1. Sustainability
We find that funds with higher sustainability ratings (i.e., more Morningstar globes) receive
larger net flows during the crisis. Figure 10 plots cumulative net fund flows across funds with
different globe ratings over the February 20 to April 30, 2020 time period. Panel A plots
aggregate cumulative flows for each globe category, which we compute as total cumulative
net flows into that category divided by that category’s total net assets on February 19, 2020.
The panel shows a near-monotonic relation between those ratings and net fund flows, with
five-globe funds having the largest flows and one-globe funds the lowest flows. Panel B
focuses on the five-globe and one-globe categories. For both of them, the panel plots average
fund-level cumulative flows across all funds in the given category, scaled as a percent of the
fund’s February 19, 2020 total net assets, along with the 95% confidence intervals. The panel
shows that five-globe funds receive significantly larger net flows than one-globe funds. In
short, investors favor sustainable funds when moving their money during the crisis.
Figure 11 shows a similar pattern based on a different measure of sustainability: an
indicator of whether the fund employs exclusions in its investment process. Funds that do
not employ exclusions, which account for the vast majority of funds, experience net outflows
during the crisis. However, funds that do employ exclusions experience net inflows, and the
difference between the two groups is statistically significant.
26C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 10. Fund Flows and Sustainability Ratings. This figure plots net fund flows over theFebruary 20 to April 30, 2020 period for categories of funds sorted by Morningstar sustainability ratings.Panel A plots aggregate cumulative net flows for each of the five globe categories. Flows are aggre-gated within each category and accumulated over time, then scaled by the category’s total net assetson February 19, 2020. Panel B plots the average across funds of cumulative net flows as a percent ofthe fund’s February 19, 2020 total net assets, for the five- and one-globe categories only. Unlike inPanel A, the sample in Panel B is restricted to funds with at least $15 million of total net assets as ofJanuary 31, 2020 and the net fund flow percentage is winsorized at the 2.5% and 97.5% levels. Panel Balso plots 95% confidence intervals, with standard errors clustered on the Morningstar Institutional Category.
27C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 11. Fund Flows and Exclusions. This figure plots net fund flows over the February 20 to April30, 2020 period for two categories of funds: those that do and do not employ exclusions in their investmentprocess. Panel A plots aggregate cumulative net flows for both categories. Flows are aggregated within eachcategory and accumulated over time, then scaled by the category’s total net assets on February 19, 2020.Panel B plots the average across funds of cumulative net flows as a percent of the fund’s February 19, 2020total net assets, for both categories. Unlike in Panel A, the sample in Panel B is restricted to funds with atleast $15 million of total net assets as of January 31, 2020 and the net fund flow percentage is winsorized atthe 2.5% and 97.5% levels. Panel B also plots 95% confidence intervals, with standard errors clustered onthe Morningstar Institutional Category.
Figure 12 unpacks sustainability into its E, S, and G dimensions. As before, we separate
funds into high-E, low-E, high-S, low-S, high-G and low-G, where the high (low) group
always denotes the top (bottom) 30% of funds. Figure 12 shows that cumulative aggregate
net flows during the crisis are larger for high-E funds than for low-E funds. Low-E funds
experience substantial outflows of 2.7% of assets, whereas the outflows from high-E funds
are only 0.3%. Net flows are also larger for high-G funds than for low-G funds: low-G funds
have outflows of 2.0% whereas high-G funds’ outflows are only 1.0%. High-S funds actually
experience larger outflows than low-S funds, but the difference between the two categories’
28C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
total flows is relatively small, only 0.7%, and the average outflow is in fact slightly larger for
the low-S category. The effect of sustainability on fund flows thus seems driven by E and, to
a lesser extent, G. Investors seem to have retained their focus on environmental issues even
during the health crisis of 2020.
Figure 12. Fund Flows and ESG Scores. This figure plots net fund flows over the February 20 to April30, 2020 period for funds in the top 30% (“high”) and bottom 30% (“low”) of environmental, social, andgovernance sustainability scores. The left panels plot aggregate cumulative net flows for the high and lowE, S, and G categories. Flows are aggregated within each category and accumulated over time, then scaledby the category’s total net assets on February 19, 2020. The right panels plot the average across funds ofcumulative net flows as a percent of the fund’s February 19, 2020 total net assets, for both the high andlow categories. Unlike in the left panels, the samples in the right panels are restricted to funds with at least$15 million of total net assets as of January 31, 2020 and the net fund flow percentage is winsorized at the2.5% and 97.5% levels.
29C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 3 reports results from cross-sectional regressions of crisis-period net fund flows on
the sustainability variables as well as style fixed effects and a large number of fund and
industry controls. The sustainability variables remain significant even with these controls.
The exclusions variable is the most robust, with t-statistics ranging from 2.50 to 4.02 across
the seven different specifications. The indicator for high-E (i.e., green) funds is also generally
significant, with t-statistics ranging from 1.89 to 3.09. The five-globe indicator is significant
when included on its own as well as in several other specifications, but it loses significance
when the exclusions indicator is included. Overall, these results indicate that investors favor
sustainable funds while reallocating money during the pandemic of 2020.
Table 3Determinants of Net Fund Flows During the Crisis
The table reports slope coefficients estimated from regressions of net fund flows in February 20 to April 30,2020 on fund characteristics and controls. A fund’s net flow is expressed as a percent of the fund’s February19, 2020 total net assets. Flows are winsorized at the 2.5% and 97.5% levels. Global category fixed effects arebased on the Morningstar Global Category variable. Fund-level controls include an indicator for a growthtilt, the log of the fund’s age in days, the log of the fund’s January 31, 2020 total net assets (TNA), turnoverratio as of January 2020, net expense ratio as of January 2020, net cash position (as a percent of TNA) as ofJanuary 2020, Morningstar medal rating as of January 2020, and market beta estimated from the October1, 2019 to January 31, 2020 period. Industry controls include the fund’s net position as a percent of TNA inbasic materials, communication services, consumer cyclical, consumer defensive, healthcare, industrials, realestate, technology, energy, financial services, and utilities. Standard errors are clustered on the MorningstarInstitutional Category. t-statistics are in brackets.
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
I(5 Sustainability Globes) 1.76 0.92 1.36 1.44 0.70 1.16 1.22[3.19] [1.82] [2.23] [2.32] [1.23] [1.77] [1.82]
I(Employs Exclusions) 2.84 2.61 2.70 2.75 2.69 2.95 3.11[3.37] [3.09] [2.50] [2.53] [4.02] [3.46] [3.43]
I(Greener E) 1.67 1.02 1.04 1.91[3.09] [2.08] [1.89] [2.72]
I(Greener S) −0.48 −0.43 −0.39 −0.88[−0.93] [−0.83] [−0.64] [−1.22]
I(Greener G) 0.76 0.63 1.00 1.26[1.39] [1.09] [1.31] [1.69]
Star Rating 1.83 1.78 1.90 1.82 1.76 1.89 1.70 1.38 1.49[7.49] [6.04] [6.00] [7.51] [6.04] [5.96] [5.77] [3.91] [4.01]
Global Category FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesFund-Level Controls No No Yes Yes No No Yes Yes No No Yes YesIndustry Controls No No No Yes No No No Yes No No No YesObservations 2,082 1,863 1,434 1,390 2,082 1,863 1,434 1,390 1,503 1,348 1,037 1,020Adjusted R2 0.02 0.08 0.11 0.11 0.02 0.09 0.11 0.11 0.01 0.08 0.10 0.11
30C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
4.2. Other Determinants of Fund Flows
Figure 13 plots cumulative net fund flows for funds with different star ratings as of January
31, 2020. Panel A shows a monotonic flow-star relation, with higher-star funds receiving
higher net flows. Panel B shows that the differences in average cumulative net flows are
statistically significant: five-star funds receive significantly larger average net flows than
three-star funds, whose average flows are significantly larger than those of one-star funds.
The positive flow-star relation is highly significant also after controlling for style fixed effects
and the other controls (see Table 3). This relation is not surprising, however, because star
ratings are constructed based on past returns, which are well known to have predictive power
for fund flows. Star ratings essentially perform the role of catch-all controls for past returns
in our regressions.
Figure 13. Fund Flows and Star Ratings. This figure plots net fund flows over the February 20 toApril 30, 2020 period for categories of funds sorted by Morningstar star ratings. Panel A plots aggregatecumulative net flows for each of the five star categories. Flows are aggregated within each category andaccumulated over time, then scaled by the category’s total net assets on February 19, 2020. Panel B plotsthe average across funds of cumulative net flows as a percent of the fund’s February 19, 2020 total net assets,for the five-, three-, and one-star categories only. Unlike in Panel A, the sample in Panel B is restricted tofunds with at least $15 million of total net assets as of January 31, 2020 and the net fund flow percentage iswinsorized at the 2.5% and 97.5% levels. Panel B also plots 95% confidence intervals, with standard errorsclustered on the Morningstar Institutional Category.
31C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 14 shows that growth funds receive significantly larger net flows than value funds.
This relation holds largely at the style level because it vanishes when we run regressions that
include style fixed effects. We include the growth indicator variable among the fund-level
controls in Table 3 but we suppress it because it is never statistically significant.
Figure 14. Fund Flows and Growth vs. Value Funds. This figure plots net fund flows over theFebruary 20 to April 30, 2020 period for growth vs value funds, as determined by the Morningstar equitystyle box. Panel A plots aggregate cumulative net flows for both categories. Flows are aggregated withineach category and accumulated over time, then scaled by the category’s total net assets on February 19,2020. Panel B plots the average across funds of cumulative net flows as a percent of the fund’s February19, 2020 total net assets, for both categories. Unlike in Panel A, the sample in Panel B is restricted tofunds with at least $15 million of total net assets as of January 31, 2020 and the net fund flow percentage iswinsorized at the 2.5% and 97.5% levels. Panel B also plots 95% confidence intervals, with standard errorsclustered on the Morningstar Institutional Category.
32C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
As in our analysis of fund performance, we also report our regression results in specifi-
cations without style fixed effects. The results are quite similar to those reported in Table
3, as we show in the Appendix. The Appendix also reports subperiod (crash and recovery)
results for Table 3, as well as for its version with no style fixed effects.
5. Conclusions
We analyze the performance and flows of U.S. active equity mutual funds during the COVID-
19 crisis. We find that most active funds underperform passive benchmarks during the crisis,
contradicting the hypothesis that active funds outperform in recessions. This underperfor-
mance is particularly large when measured against the S&P 500 index, but it is observed also
relative to fund-specific style benchmarks. Funds with high sustainability ratings and high
star ratings outperform those with low sustainability ratings and low star ratings, respec-
tively. When reallocating capital across funds, investors favor funds with high sustainability
ratings and funds that apply exclusion criteria. That investors retain their focus on sustain-
ability during a major crisis indicates that sustainability is not just a luxury good.
While this paper appears to be the first to analyze the performance and flows of active
funds during the COVID-19 crisis, it leaves plenty of room for future research. For example,
while we focus on equity funds, fixed income funds also deserve a careful investigation. So
do the sources of active funds’ underperformance during the crisis. It would also be useful
to extend our work to deepen our understanding of the dependence of investors’ demand for
sustainability on economic conditions.
33C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
REFERENCES
Albuquerque, Rui A., Yrjo Koskinen, Shuai Yang, and Chendi Zhang, 2020, Love in the timeof COVID-19: The resiliency of environmental and social stocks, Working paper, BostonCollege.
Alfaro, Laura, Anusha Chari, Andrew N. Greenland, and Peter K. Schott, 2020, Aggregateand firm-level stock returns during pandemics, in real time, Covid Economics 4, 2–24.
Barber, Brad M., Xing Huang, and Terrance Odean, 2016, Which factors matter to investors?Evidence from mutual fund flows, The Review of Financial Studies 29, 2600–2642.
Baumol, William J. and Wallace E. Oates, 1979, Economics, Environmental Policy, andQuality of Life, Inglewood Cliffs, NJ: Prentice-Hall.
Bialkowski, Jedrzej, and Laura T. Starks, 2016, SRI funds: Investor demand, exogenousshocks and ESG profiles, Working paper.
Bretscher, Lorenzo, Alex Hsu, and Andrea Tamoni, 2020, The supply channel of uncertaintyshocks and the cross-section of returns: Evidence from the COVID-19 crisis, Workingpaper, LBS.
Carhart, Mark M., 1997, On persistence in mutual fund performance, Journal of Finance52, 57–82.
Ding, Wenzhi, Ross Levine, Chen Lin, and Wensi Xie, 2020, Corporate immunity to theCOVID-19 pandemic, Working paper, University of Hong Kong.
Elton, Edwin J., Martin Jay Gruber, Sanjiv Das, and Matthew Hlavka, 1993, Efficiency withcostly information: A reinterpretation of evidence from managed portfolios, Review ofFinancial Studies 6, 1–22.
Elton, Edwin J., Martin Jay Gruber, and Christopher R. Blake, 2001, A first look at theaccuracy of the CRSP mutual fund database and a comparison of the CRSP and Morn-ingstar mutual fund databases, Journal of Finance 56.
Evans, Richard, 2010, Mutual fund incubation, Journal of Finance 65, 1581–1611.
Fama, Eugene F., and Kenneth R. French, 1993, Common risk factors in the returns onstocks and bonds, Journal of Financial Economics 33, 3–56.
Fama, Eugene F., and Kenneth R. French, 2010, Luck versus skill in the cross section ofmutual fund returns, Journal of Finance 65, 1915–1947.
Fama, Eugene F., and Kenneth R. French, 2015, A five-factor asset pricing model, Journalof Financial Economics 116, 1–22.
34C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Fahlenbrach, Rudiger, Kevin Rageth, and Rene M. Stulz, 2020, How valuable is financialflexibility when revenue stops? Evidence from the Covid-19 crisis, Working paper, SwissFinance Institute.
Ferson, Wayne E., and Rudi W. Schadt, 1996, Measuring fund strategy and performance inchanging economic conditions, Journal of Finance 51, 425–461.
Gerding, Felix, Thorsten Martin, and Florian Nagler, 2020, The value of fiscal capacity inthe face of a rare disaster, Working paper, Bocconi University.
Glode, Vincent, 2011, Why mutual funds “underperform”, Journal of Financial Economics99, 546–559.
Gormsen, Niels Joachim, and Ralph SJ Koijen, 2020, Coronavirus: Impact on stock pricesand growth expectations, Working paper, University of Chicago.
Gruber, Martin J., 1996, Another puzzle: The growth in actively managed mutual funds,Journal of Finance 51, 783–810.
Haddad, Valentin, Alan Moreira, and Tyler Muir, 2020, When selling becomes viral: Dis-ruptions in debt markets in the COVID-19 crisis and the Fed’s response, Working paper,UCLA.
Hartzmark, Samuel M., and Abigail B. Sussman, 2019, Do investors value sustainability? Anatural experiment examining ranking and fund flows, Journal of Finance 74, 2789–2837.
Jensen, Michael C., 1968, The performance of mutual funds in the period 1945–1964, Journalof Finance 23, 389–416.
Kacperczyk, Marcin, Stijn Van Nieuwerburgh, and Laura Veldkamp, 2014, Time-varyingfund manager skill, Journal of Finance 69, 1455–1484.
Kacperczyk, Marcin, Stijn Van Nieuwerburgh, and Laura Veldkamp, 2016, A rational theoryof mutual funds’ attention allocation, Econometrica 84, 571–626.
Kargar, Mahyar, Benjamin T. Lester, David Lindsay, Shuo Liu, and Pierre-Oliver Weill,2020, Corporate bond liquidity during the COVID-19 crisis, Working paper, Universityof Illinois.
Kosowski, Robert, 2011, Do mutual funds perform when it matters most to investors? USmutual fund performance and risk in recessions and expansions, Quarterly Journal ofFinance 1, 607–664.
Lins, Karl V., Henri Servaes, and Ane Tamayo, 2017, Social capital, trust, and firm perfor-mance: The value of corporate social responsibility during the financial crisis, Journal ofFinance 72, 1785–1824.
35C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Malkiel, Burton G., 1995, Returns from investing in equity mutual funds 1971 to 1991,Journal of Finance 50, 549–572.
Moskowitz, Tobias J., 2000, Mutual fund performance: An empirical decomposition intostock-picking talent, style, transactions costs, and expenses: Discussion, Journal of Fi-nance 55, 1695–1703.
Nofsinger, John, and Abhishek Varma, 2014, Socially responsible funds and market crises,Journal of Banking and Finance 48, 180–193.
O’Hara, Maureen, and Xing Alex Zhou, 2020, Anatomy of a liquidity crisis: Corporate bondsin the COVID-19 crisis, Working paper, Cornell University.
Pagano, Marco, Christian Wagner, and Josef Zechner, 2020, Disaster resilience and assetprices, Covid Economics 21, 1–39.
Pastor, Lubos, and Robert F. Stambaugh, 2002, Mutual fund performance and seeminglyunrelated assets, Journal of Financial Economics 63, 315–349.
Pastor, Lubos, and Robert F. Stambaugh, 2012, On the size of the active managementindustry, Journal of Political Economy 120, 740–781.
Pastor, Lubos, Robert F. Stambaugh, and Lucian A. Taylor, 2015, Scale and skill in activemanagement, Journal of Financial Economics 116, 23–45.
Pastor, Lubos, Robert F. Stambaugh, and Lucian A. Taylor, 2017, Do funds make morewhen they trade more?, Journal of Finance 72, 1483–1528.
Pastor, Lubos, Robert F. Stambaugh, and Lucian A. Taylor, 2020, Sustainable investing inequilibrium, Journal of Financial Economics, forthcoming.
Ramelli, Stefano, and Alexander F. Wagner, 2020, Feverish stock price reactions to COVID-19, Working paper, University of Zurich.
Schrimpf, Andreas, Hyun Song Shin, and Vladyslav Sushko, 2020, Leverage and marginspirals in fixed income markets during the COVID-19 crisis, Working paper, BIS.
Sensoy, Berk, 2009, Performance evaluation and self-designated benchmark indexes in themutual fund industry, Journal of Financial Economics 92, 25–39.
Wermers, Russ, 2000, Mutual fund performance: An empirical decomposition into stock-picking talent, style, transactions costs, and expenses, Journal of Finance 55, 1655–1695.
36C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
1-3
6
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Covid Economics Issue 38, 16 July 2020
Copyright: Thorsten Beck, Burton Flynn and Mikael Homanen
COVID-19 in emerging markets: Firm-survey evidence1
Thorsten Beck,2 Burton Flynn3 and Mikael Homanen4
Date submitted: 10 July 2020; Date accepted: 10 July 2020
Using survey responses across nearly 500 listed firms in 10 emerging markets from early April, we find the vast majority of firms were negatively affected by COVID-19. Firms reacted by reducing investment rather than payroll. There is a surprising degree of support vis-à-vis employees, customers, other stakeholders and broader society. Although stock prices initially reacted to the impact of the crisis, delayed stock price reactions suggest evidence of inefficient markets. Furthermore, we find evidence that stakeholder-centric firms experienced lower stock price declines during the crisis drawdown.
1 We are grateful for excellent assistance by Rabi Uddin Nasir, Michael Genio Aguila, Ivan Nechunaev and Ara Reyes Tisalona, and comments from participants at the Cass PhD workshop and RCN Seminar series. The views expressed in this paper are the authors’ only and do not necessarily represent those of the PRI.
2 City’s Business School, University of London, and CEPR3 Terra Nova Capital, Evli Group and City’s Business School, University of London.4 PRI and City’s Business School, University of London.
37C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
1. Introduction
The economic fallout from the COVID-19 pandemic has been severe as reflected in recently
released GDP numbers for the first quarter of 2020 and predictions for the rest of the year
(IMF, 2020). Whole industries have been shut and firm-level evidence confirms the
aggregate number in terms of lost jobs, output and investment. Most of the evidence so far,
however, has been for advanced countries. This paper presents firm-level evidence on the
impact of COVID-19 for 10 emerging markets, combining survey answers with financial
statement information and stock price returns.
Many advanced countries have reacted to the pandemic with lockdown policies that
brought a large part of the economy to a standstill, including industries that rely on direct
customer contact but also manufacturing and construction in some countries. Among
emerging markets, there has been variation in responses, with some countries going for
strict lockdowns, such as Vietnam, while others keeping their economies relatively open,
such as Turkey. These differences in timing and severity of lockdown policies have also
implications for firms and their employees. More generally, higher informality, fewer jobs
that can be done from home, and more limited state capacity make both public-health
oriented containment and their enforcement less effective, while limited fiscal space and
limited access to international financial markets make economic support policies more
difficult to implement (Djankov and Panizza, 2020). The trade-off between public health and
economic survival is therefore more biased towards the latter in many emerging markets
(Alon et al., 2020; Hevia and Neumeyer, 2020).
We present survey evidence on 488 firms across 10 emerging markets in Europe, Africa, the
Middle East, and Asia on the impact of COVID-19, firms’ financial and investment reaction
and their approach towards employees, other stakeholders and the broader society.
Specifically, we explore cross-country and cross-sector variation in (i) the effect of the
pandemic on firms, (ii) firms’ reaction in investment and labour demand and (iii) firms’
reactions in their relationships with employees, suppliers and other stakeholders. We also
test whether the effect of the crisis on firms and their reaction is reflected in stock returns.
38C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
As the pandemic started in China and then spread to Europe and North America, many
emerging markets experienced first an economic and then a public health shock.
Specifically, disruptions in supply chains and drop in international demand were often the
first negative signals before the public health crisis reached many of these countries. This
gave both governments and firms (the focus of this study) time to react to the pandemic.
However, firm-level analysis suggests that many firms in developing countries have limited
liquidity to survive a longer economic lockdown that deprives them of their revenue
stream.1 Furthermore, integration into global supply chains makes many firms more
vulnerable during global recessions (Accetturo and Giunta 2019), while procyclical bank
behaviour might result in lower access to external funding when it is needed most.
Firms can react in different ways to such an unprecedented shock. On the one hand,
investment plans can be cancelled rather quickly; such a reaction might reflect firms’
assumption that there will not be a V-shaped quick recovery but rather a drawn-out crisis or
loss of access to external funding. On the other hand, many emerging markets’ labour
markets are often considered more flexible than those of advanced countries, so that quick
reductions of payrolls are another option to reduce costs. However, relationships between
firms and their stakeholders, such as employees, customers, suppliers and society at large,
are often based not only on contractual but also informal and personal links, so that firms
can react quickly and reduce commitment but might also have to worry about long-term
implications of undermining important relationships with business partners and
stakeholders.
Our survey results for almost 500 listed firms across 10 emerging markets show at least
three in four firms being negatively impacted by the pandemic. Surprisingly few firms,
however, expect to breach their covenants or see a need for raising additional capital.
About half of the firms have received or expect to receive government support. Firms have
reacted primarily reducing investment spending and much less through layoffs. Meanwhile,
1 Bosio et al. (2020a) find that in high- and middle-income countries, without revenues, the average firms have liquidity to survive between 12 and 38 weeks, while Bosio (2020b) find for firms in lower and lower-middle income countries, the average survival time across industries ranges from 6 to 28 week and 6 and 18 weeks once collapsed export demand is taken into account.
39C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
some firms cut back on executive compensation, and more firms expanded employee
benefits than cut them. The large majority of firms have acted before their governments
imposed measures, shown flexibility vis-à-vis customers other stakeholders, and provided
donations to support society at large or have shifted business operations to fulfil pandemic
needs. This shows a picture of firms focusing on short-term needs of stakeholders,
protecting labour and long-term relationships, thus a picture not consistent with short-term
focused shareholder maximisation value, but longer-term value maximisation for both share
and other stakeholders.
A second major finding is that there is significant variation across countries, sectors and
firms in both impact and reaction. While sector-variation matters most for variation in the
impact of the crisis and government support, country-variation seems to matter most for
variation in firm reaction to the crisis. Interestingly, variation in whether firms reacted to
the pandemic before government-imposed actions varies more with firm-level
characteristics than across countries or sectors.
Third, delayed share price reactions to the impact of the crisis and firms’ reactions suggest
inefficient markets in which astute investors were able to capture excess returns. While
there was a reaction between late February and late April, firms that stated in early April
that they were negatively affected by the pandemic saw a 6.1% lower cumulative return
between late April and late June. Additionally, firms that claimed to be positively affected
saw a 18.3% higher return in the following two months though this result is not statistically
significant owing to the small sample size (only 13 such firms). Furthermore, firms that
stated in April that they were reducing investment saw a lower return both between late
April and late June. These delayed reactions to publicly available information are not
consistent with the semi-strong form of efficient market hypothesis.
Fourth, we also document that more stakeholder-centric firms (e.g. those which did not
reduce employee benefits, took measures to protect stakeholders or made donations to
fight the pandemic) experienced lower stock price declines and did not significantly
underperform during the late April to late June period, suggesting that the financial markets
valued these stakeholder-centric corporations more than their counterparts during the
crisis.
40C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Our paper is part of a small but rapidly expanding literature exploring the impact of the
pandemic. While several papers have used stock market data to explore the real sector
implications of the COVID-19 recession (e.g., Alfaro et al., 2020 and Capelle-Blancard and
Desrozier, 2020) and variation of stock market returns across firms (e.g. Ding et al., 2020,
and Ramelli and Wagner, 2020), there are few papers using survey data capturing firms’
reactions to the COVID-19 crisis and those few primarily focus on advanced rather than
developing countries (e.g., Bennedsen et al., 2020 for Denmark and Li et al., 2020, for the
US). While there are household surveys in developing countries (e.g., Gomez et al., 2020 and
Rahman et al., 2020), our paper relies on (one of) the first firm-level surveys across
developing countries.
Before proceeding, we would like to state a few caveats. First, the firms in our survey are
medium sized to large, formal and listed. The survey does not provide any insights about the
reaction of smaller and/or informal firms. It is notable, however, that together the operating
income of the surveyed firms make up between 0.5% and 8% of the respective GDP.2 More
importantly, Alfaro et al. (2019) show that consistent with Gabaix (2011) large firms are
systemically important in emerging markets, as idiosyncratic shocks to large firms
significantly correlate with GDP growth. Second, survey responses might be biased, though
their use has become increasingly popular in corporate finance, with evidence showing that
self-reporting is significantly correlated with publicly and verifiable information. Finally, our
survey provides insights at a specific point in time in early April, in what in most countries
was the early stage of the pandemic and government responses to it. However, we did
follow up with an additional survey in early June asking for an update on prior survey
responses and found that little had changed.
The remainder of the paper is structured as follows. Section 2 discusses our data collection
process. Section 3 presents some preliminary analysis, while section 4 provides more formal
regression analysis. Section 5 shows the relationship between the impact of and reaction of
2 As we do not have value-added data for individual firms, we estimate the share of the firms in the economy by dividing the sum of 2018 operating income of all sample firms in USD by 2018 GDP in USD of their respective countries. The ratio varies between 0.2% for Turkey to 7.6% for South Africa.
41C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
firms to the pandemic and buy-and-hold returns. Section 6 concludes.
2. Data
In early April 2020, we sent out a short 13-question survey to 630 firms across 10 emerging
markets, including Bangladesh, Indonesia, Malaysia, Pakistan, Philippines, Saudi Arabia,
South Africa, Thailand, Turkey and Vietnam.3 All our firms are listed on stock exchanges and
are active across all sectors of the economy.
Out of the 630 firms, 488 responded, 98% of them between April 2 and 24, thus a response
rate of 78%.4 In early June, we followed up with these 488 firms which answered the first
set of questions with additional questions and inquiring whether their responses to the first
survey questions had changed. 415 firms of the 488 firms responded, thus a response rate
of 85%. Appendix A1 lists the survey questions and Appendix Table A2 lists the exact split of
responses.
We augment the survey data with financial statement and market information on all firms,
specifically, sales growth over the past year to capture the firm’s growth path, Tobin’s q
(market value / book value of equity) to proxy for growth opportunities, log of total assets in
million USD to capture the size of the firm, cash divided by total assets to gauge the liquidity
position, capital expenditures divided by total assets as investment indicator and the
leverage ratio as measured by debt-to-assets ratio. We report descriptive statistics in Table
1. In regression analysis, we winsorise sales growth, cash divided by total assets and capital
expenditures divided by total assets at the 1st and 99th percentiles to avoid the impact of
outliers.
3 Between June 2019 and March 2020, the portfolio managers of a Finnish-based emerging markets fund requested one-on-one private meetings with the CEOs of 1,476 firms, comprised of all listed companies with greater than $250,000 of average daily trading volume across these ten countries. Various executives (mainly CEOs, CFOs, and IR personnel) at 630 of the firms accepted the meeting requests. In April 2020, the fund managers emailed the executives they met at all 630 firms with 13 questions regarding how the coronavirus pandemic is affecting their firm and how their firm is responding. 4 91% of responses were received by Apr 17; the last response was received on 16 June.
42C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 1: Summary Statistics Panel A: April Survey - operational and governance variables
(1) (2) (3) (4) (5) (6) (7) (8) VARIABLES N mean sd min max p25 p50 p75 Have your operations been affected by the pandemic? (negatively) 439 0.697 0.460 0 1 0 1 1 Have you or might you breach covenants? 439 0.0524 0.223 0 1 0 0 0 Would you need to raise capital to weather the pandemic? 439 0.194 0.396 0 1 0 0 0 Do you expect to receive government support? 439 0.358 0.480 0 1 0 0 1
(1) (2) (3) (4) (5) (6) (7) (8)
VARIABLES N mean sd min max p25 p50 p75 Have you reduced investment spending? 439 0.579 0.494 0 1 0 1 1 Have you made any layoffs? 439 0.0820 0.275 0 1 0 0 0 Made any changes to employee benefits? (negatively) 416 0.0721 0.259 0 1 0 0 0 Have you made changes to the dividend or share buyback plan? 439 0.187 0.390 0 1 0 0 0 Have you made changes to executive compensation plans? 439 0.175 0.381 0 1 0 0 0
Panel B: April survey – stakeholder variables
(1) (2) (3) (4) (5) (6) (7) (8) VARIABLES N mean sd min max p25 p50 p75 Have you taken any measures to protect your stakeholders? 439 0.911 0.285 0 1 1 1 1 Continued to pay employees or services providers for disrupted services?
439 439 0.465 0.499 0 1 0 0 1
Provided financial flexibility to any customers or business partners? 439 0.392 0.489 0 1 0 0 1 Made any donations to help fight the pandemic? 439 0.781 0.414 0 1 1 1 1
Panel C: June survey – additional variables
(1) (2) (3) (4) (5) (6) (7) (8) VARIABLES N mean sd min max p25 p50 p75 Did you take action before the government? 267 0.700 0.459 0 1 0 1 1 Did your revenue decline in the last two months? 208 0.856 0.352 0 1 1 1 1 How long before you expect to recover to the pre-pandemic revenue? 137 2.620 0.979 1 5 2 2 3
Panel D: Firm controls
(1) (2) (3) (4) (5) (6) (7) (8) VARIABLES N mean sd min max p25 p50 p75 Sales Growth 439 1.351 3.495 0.0181 57.20 0.326 0.692 1.345 Leverage 439 25.73 18.76 0 93.02 8.845 25.53 38.69 Total Assets 439 5,701 14,923 4.536 87,503 285.7 875.0 3,539 Cash / Total Assets 439 8.308 8.908 0.0553 61.60 2.705 5.707 10.71 Capex / Total Assets 439 4.488 4.866 0.0029 25.84 0.647 3.150 6.706 Sales Growth 439 0.102 0.335 -0.661 3.777 -0.027 0.0673 0.170 Notes: The variable “How long before you expect to recover to the pre-pandemic revenue?” takes a value between 1 and 5. The variable is equal to one if companies expect recovery in the first half of 2020, two for second half of 2020, three for first half of 2021 and so forth. In total 4.5% of the firms responded with 1, 56% with 2, 25% with 3, 5% with 4 and 8% with 5, out of a total of 156 responses that provided full details of recovery expectations (not adjusting for whether firm controls are available). The variable "Made any changes to employee benefits? (negatively)" has less than 439 observations, because whether benefits were negatively or positively changed was extrapolated from qualitative answers provided by the corporate responses. In few cases, it was not possible to evaluate whether they were positive, negative or affected at all. Firm controls were collected from the S&P Capital IQ Platform. Total Assets, Cash / Total Assets, Capex / Total Assets and Sales Growth are all winsorized at the 1st and 99th percentile.
43C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Further, we use share prices of these firms to compute buy-and-hold returns between
February 24 and April 24, and between April 24 and June 24. The sample using stock prices
is smaller than our overall sample, as we do not have stock price data for Bangladeshi firms
– the Dhaka Stock Exchange was closed for two months beginning on 26 March. All financial
statement and stock price data were retrieved from S&P Capital IQ.
We formally test whether there is a selection bias in survey responses, i.e., whether the
survey respondents differ significantly from the non-respondents. Specifically, in Appendix
Table A3 we test for differences across the firm-level variables discussed above between
these two samples and do not find any significant difference across firm characteristics
except that non-respondents had higher sales growth in the previous years. This assures us
that the survey responses provide an adequate picture of firms in this segment. We also test
for the differences across the firm-level variables between firms that responded to both
April and June surveys and those that responded only to the April survey and find again
mostly insignificant differences, except that pre-crisis capital expenditures are higher for
firms responding only to the survey in April.
The survey questions are divided into three blocks. In the first block, we asked firms
whether their operations were affected by the pandemic, they might breach covenants,
they need to raise capital and they expect to receive government support. A second block
considered adjustment in firms’ operation and plans; whether they have reduced
investment, laid off staff, made changes to employee benefits, made changes to executive
compensation and made changes to the dividends or share buyback plans. Finally, we asked
questions on how firms dealt with business partners and society more generally; whether
they voluntarily have taken any measures to protect their employees and other
stakeholders, continued to pay employees or service providers for disrupted services,
provided financial flexibility to any customers or business partners and made any donations
to help fight the pandemic or shifted business operations to fulfil pandemic needs. In
addition to yes/no answers, most firms gave more details on some of these questions,
which we exploit below.
In June, we asked firms whether there had been any updates on their responses to the April
survey, plus three additional questions. First, by how much approximately their sales had
44C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
fallen over the past two months (April and May). Second, by when they expected a recovery
to the pre-crisis level of revenues. Third, whether they adjusted their operations to the
circumstances of COVID-19 before government measures came into place or following the
issue of such measures.
3. COVID-19 in emerging markets
The survey responses from April and June show clearly an overwhelmingly negative impact
of the pandemic on firms, a quick reaction of firms to the pandemic and quite some
flexibility in terms of their relationships with stakeholders. However, we also find significant
variation across countries and across sectors. Table 1 lists the descriptive statistics of the
different variables. Figure 1 shows the variation for selected survey variables across
countries and sectors, and Appendix Tables A4 and A5 show the variation for all the survey
variables across countries and sectors.
Figure 1: COVID-19 across countries and sectors
72%
89%
57%
81%89%
71%65% 68% 68% 62%
68%
2% 1% 5% 2% 1% 4% 3% 8%
All
Bangladesh
Indonesia
Malaysia
Phillipines
Pakistan
Saudia Arabia
Thailand
Turkey
Vietnam
South Africa
Have you been affected by the pandemic?
Negative Positive
61%
86%
59%68%
53%65%
83%76%
68%
86%
42%
17% 3% 10% 1% 6%
Comm
unication
Consumer…
Consumer…
Energy
Financials
HealthCare
Industrials
IT Materials
RealEstate
Utilities
Have you been affected by the pandemic?
Negative Positive
45C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
The vast majority of surveyed firms have been negatively affected by the COVID-19 crisis.
Specifically, in April 2020, 69% of firms report a negative impact and less than 3% a positive
impact, with 15% reporting no impact and 5% not answering the questions. The remaining
11% of firms that answered the question reported an impact but did not specify the type of
impact. However, the effect varies a lot across countries and sectors; ranging from a
negative impact to 87% of firms in the Philippines to 55% in Indonesia, and between 85% in
real estate and 43% in utilities (Figure 1). The sectoral variation is not surprising, as firms in
essential sectors (e.g., utilities) were hit less and firms in less essential sectors (e.g.,
consumer discretionary) and sectors relying heavily on physical customer interaction (e.g.,
real estate) were hit most.
51%67% 65%
48%66%
20% 24%
64%
37%50%
72%
5%
7% 4%
4%
4%
13% 5%
2%
4%
2%
2%56%
74% 69%
51%
70%
33% 29%
66%
41%52%
74%
All
Bangladesh
Indonesia
Malaysia
Phillipines
Pakistan
Saudia Arabia
Thailand
Turkey
Vietnam
South Africa
Have you reduced investment spending in response to the
pandemic?
Yes Unanswered
38%
66%51%
67%41%
56% 57% 53% 63%45%
5%
4%8%
1%
3%
3% 1% 5%2%
8%43%
70%59%
68%
44%53%
59% 58% 58%65%
54%
Comm
unication
Consumer…
Consumer Staples
Energy
Financials
HealthCare
Industrials
IT Materials
RealEstate
Utilities
Have you reduced investment spending in response to the
pandemic?
Yes Unanswered
50% 46% 52%
77%61%
37% 28% 33%46%
83%3% 4%
2%
5%
3%
2%2%
2%
3%
6%
53%
37%
50% 54%
82%
63%
38%30%
36%
49%
90%
All
Bangladesh
Indonesia
Malaysia
Phillipines
Pakistan
Saudia Arabia
Thailand
Turkey
Vietnam
South Africa
Have you continued to pay employees/service providers if
the service was disrupted?
Yes Unanswered
54% 46% 39%56%
38% 47%34%
64% 63%
8%
5%3%
3%
2%
2%3%
2%
6% 4%
1%
48%60%
50%42%
58%
40%50%
37%
70% 66%
9%
Comm
unication
Consumer…
Consumer Staples
Energy
Financials
HealthCare
Industrials
IT Materials
RealEstate
Utilities
Continued to pay employees or services providers for disrupted
services?
Yes Unanswered
46C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
In early June, we followed up with a question of how much revenues had fallen in April and
May. While most firms did not give a specific number, 50% of firms indicated a fall and 28%
declined to answer, thus confirming their response from April that they were negatively
impacted by the pandemic. Meanwhile, 13% reported minimal impact and 9% said revenues
actually increased.
In June, we also asked firms by when they expected a recovery to pre-crisis revenues. There
is wide variation in responses to this question, ranging from Q2 2020 to sometime in 2023:
11% of firms did not see any significant deterioration, consistent with responses above. 24%
of firms expect a recovery in 2020, 12% expect a recovery in 2021 and 3% expect a recover
in 2022 or later, while 49% did not answer the question. The fact that half of firms did not
answer the question points to the very high uncertainty that firms across the globe,
including in emerging markets, currently operate under.
Few firms responded affirmatively to the question on whether they have breached or
expect to breach any covenants. While only six firms (1.2%) responded affirmatively in April,
93% responded negatively. If we consider the share of firms responding affirmatively or not
responding as the upper bound for the likelihood that firms breach their covenants, we find
that up to 6% might be breaching their covenants. Six additional firms reported breach of
covenants in the June survey round (a total of 2.4%).
Similarly, few firms expect to need to raise capital in the near future because of the
pandemic. Specifically, 12% (15%) responded affirmatively in April (June), while 80%
responded negatively. However, there is quite some variation across countries, with 45% of
firms in Bangladesh stating that they need to raise capital, but only 1% in South Africa, as
well as across sectors, from 28% in energy to 5% in finance.
On the other hand, more than one third of firms (35%) expected to receive government
support, ranging from 49% in Saudi Arabia to 19% in South Africa, and from 54% in
consumer discretionary goods to none in utilities. This share only increased to 37% in June
at the time of our second survey. There is also quite some variation in the form of
government support firms have received or expect to receive, with tax or social security
support relief and wage and employee benefits being listed most often, followed by debt
47C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
relief. Support with rental or utility payments or other fees and relaxation of regulations are
named less.
In terms of firm reactions to the pandemic, half of surveyed firms (50%) have reduced their
investment spending, ranging from 65% in Thailand to 16% in Pakistan, 67% in energy and
real estate to 35% in communications and financials when asked in April (Figure 1). This
only increased to 52% during our second survey in June. On the other hand, very few firms
(3%) have laid off employees – even taking into account non-responses only gets us to 9%,
suggesting that the cuts fell primarily on capital rather than wage spending. It is more, 12%
of firms have expanded employee benefits and only 7% have reduced them. Meanwhile,
17% of firms have reduced executive compensation.
In addition to cutting expenditures, another way to strengthen firms’ financial position is to
reduce/suspend dividends or stop share buy-backs. Across the ten countries, 13% of all
firms planned to do so, ranging from 17% in South Africa to none in the Pakistan, from 20%
in energy and industrials to 2% in consumer staple goods. The proportion of firms reducing
or suspending dividends increased to 18% in June.
Most of the surveyed firms have taken actions to support their stakeholders and the
broader society. 91% of the firms have taken measures to protect employees and
stakeholders beyond government-imposed regulations and 47% have continued to pay
employees and vendors even if service was disrupted. 100% of firms in Bangladesh and
Pakistan reported taking voluntary measures to protect employees and other stakeholders
while only 79% in Turkey did so, and 100% of firms in the IT and utilities sectors compared
to only 85% of consumer staples companies. 71% of firms in South Africa continued paying
employees for disrupted services compared to only 28% of firms in Thailand, and 60% of
firms in the real estate sector voluntarily paid employees while only 21% of utilities firms did
the same. 41% of surveyed firms have provided financial flexibility to customers or business
partners, ranging from 70% of firms in the Philippines to 15% in Bangladesh, and from 62%
of financials to 18% of health care.
Nearly four in five surveyed firms (77%) have made donations to support the broader
society in their community and country or shifted business operations to fulfil pandemic
48C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
needs. Donations have taken on the form of cash and PPE and other medical or hygienic
products, as well as food. To provide some examples: 26 firms shifted business operations
to manufacture personal protective equipment or sanitizing/cleaning items, while 51 firms
provided various services free of charge, such as servicing of ambulances by an auto
manufacturer, development of a donations platforms by tech companies, provision of hotel
rooms to medical workers to stay during curfew period, quarantine facilities by several hotel
groups as well as a software company campus which hosts over 2,000 people, care for
COVID-19 patients by a hospital group seeking only to recover costs from the government,
free storage of coronavirus-related supplies for NGOs fighting the virus by a self-storage
company, potential construction/conversion for temporary hospitals by a construction firm,
provision of bus fleet by a taxi company for drop off and pick up services for doctors and
nurses, assistance to government of citizen repatriation by an airline and free food packs for
medical workers, police force, military personnel and logistics drivers by a gas station
operator.
In the June follow-up survey, we also asked firms whether they adjusted their operations to
the circumstances of COVID-19 before government guidelines came into place or following
the issue of such guidelines. The large majority of firms (64%) reported that they adjusted
operations before government guidelines came into effect while 27% reported that they
waited until government measures were implemented before reacting. This may be due to
the fact that companies in emerging markets generally do not rely on their government to
lead; as one broker in Mexico put it, “companies, business men and women, do not count at
all in our government, so we need to organize ourselves for this new era.”5 However, there
is again variation across countries and sectors. For example, 72% of firms in Thailand took
proactive action whereas only 33% of firms in Turkey adjusted operations before
government measures, and 83% of firms in the utilities sector made adjustments prior to
the government’s guidelines while only 45% of energy firms did so. It is interesting to note
(and we will return to this below) that country-variation in such action does not vary with
how quickly countries imposed lockdowns.
5 Private message to one of the paper’s authors on May 5, 2020.
49C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
4. Explaining variation in COVID-19 impact and responses
While the previous section explored simple correlations, we now offer a more formal
regression analysis to explore which (i) firm characteristics, (ii) country factors and (iii)
sector factors can explain variation in the impact of COVID-19 and firm-reaction to the
pandemic. While our dependent variables are binary, we use linear regression models to
avoid the incidental parameter problem given that we saturate our models with country-
and sector fixed effects. Specifically, we run the following regression:
Yijk = a S Ik + b S Cj + g Xi + eijk (1)
Where the subscripts i, j, and k denote firm, country and sector respectively. As discussed
above, we winsorise several of the firm variables at 1st and 99th percentile. The omitted
country is Bangladesh and the omitted sector is communication, so that the other country
(sector) fixed effects denote differences in the respective country (sector) to Bangladesh
(communication). We report heteroskedasticity consistent error terms.
Before presenting regression results, we assess whether country, sector or firm
characteristics are most important in explaining variation in the different survey responses.
Specifically, Appendix Table A6 shows regressions with (i) firm-level variables, (ii) country
dummies, (iii) sectoral dummies and (iv) all three groups, for three of our survey variables.
Overall, country-variation seems to matter most for variation in adjustments in investment
spending and donations, while variation in government support varies most across sectors.
For other survey questions (not reported), firm-level variables seem to matter most for
breaching (or not) covenants, whether firms took actions before the government issued
guidance, and expected time to recovery, sector-variation matters most for the negative
impact of the crisis and changes to executive compensation, while for other variables it is
either country variation that matters or there is a mix of firm-, sector- and country-factors
that matter most.
The results in Table 2 show that larger firms are less likely to raise capital to weather the
pandemic, more likely to reduce investment spending, provide financial flexibility to
customers and make donations or shift operations to fill pandemic needs. Larger firms also
50C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 2: COVID-19 Corporate Impact and Responses Panel A
(1) (2) (3) (4)
VARIABLES
Have you been
affected by the
pandemic? (negatively)
Have you or might you breach covenants?
Would you need to raise capital to
weather the pandemic?
Do you expect to receive
government support?
Tobin’s q -0.008* 0.010 0.006 0.001 (0.005) (0.006) (0.005) (0.006) Log of Total Assets 0.011 0.008 -0.029** 0.003 (0.015) (0.006) (0.013) (0.014) Cash / Total Assets -0.001 -0.000 -0.007*** -0.002 (0.003) (0.001) (0.002) (0.003) Capex / Total Assets 0.005 -0.003 -0.004 0.005 (0.005) (0.002) (0.005) (0.006) Leverage 0.000 0.002*** 0.000 0.002* (0.001) (0.001) (0.001) (0.001) Sales Growth 0.099* 0.001 0.088 -0.117** (0.055) (0.018) (0.058) (0.053) Have you been affected by the -0.051* 0.022 0.057 pandemic? (negatively) (0.026) (0.041) (0.051) Observations 439 439 439 439 R-squared 0.143 0.102 0.160 0.118
Panel B
(1) (2) (3) (4) (5)
VARIABLES
Have you reduced
investment spending?
Have you made
changes to the dividend
or share buyback
plan?
Have you made
changes to exec com
plans?
Have you made any layoffs?
Made any negative
changes to employee benefits?
Tobin’s q 0.011*** 0.017*** 0.002 0.008 -0.005 (0.003) (0.004) (0.007) (0.006) (0.004) Log of Total Assets 0.046*** 0.007 -0.001 -0.003 -0.004 (0.014) (0.012) (0.011) (0.008) (0.007) Cash / Total Assets -0.002 -0.000 0.000 -0.000 -0.000 (0.003) (0.002) (0.002) (0.001) (0.001) Capex / Total Assets -0.003 -0.003 -0.001 0.005 -0.003 (0.005) (0.004) (0.004) (0.004) (0.003) Leverage 0.002 0.001 0.002** 0.001 0.000 (0.001) (0.001) (0.001) (0.001) (0.001) Sales Growth 0.047 -0.014 -0.018 -0.008 0.012 (0.056) (0.054) (0.043) (0.031) (0.030) Have you been affected by the 0.143*** 0.031 0.076** -0.021 0.061*** pandemic? (negatively) (0.054) (0.040) (0.036) (0.031) (0.023) Observations 439 439 439 439 416 R-squared 0.18 0.121 0.122 0.090 0.101
51C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Panel C
(1) (2) (3) (4)
VARIABLES
Have you taken any
measures to protect your
stakeholders?
Continued to pay
employees or services providers
for disrupted services?
Provided financial
flexibility to any
customers or business partners?
Made any donations
to help fight the
pandemic?
Tobin’s q 0.002 -0.007** 0.004 -0.013*** (0.002) (0.003) (0.005) (0.003) Log of Total Assets 0.011 0.014 0.046*** 0.050*** (0.010) (0.015) (0.015) (0.013) Cash / Total Assets 0.003* 0.000 -0.005* 0.001 (0.002) (0.003) (0.003) (0.003) Capex / Total Assets 0.002 -0.001 -0.005 0.001 (0.003) (0.006) (0.005) (0.005) Leverage -0.000 -0.003*** 0.001 -0.002* (0.001) (0.001) (0.001) (0.001) Sales Growth 0.002 0.128** 0.030 -0.059 (0.033) (0.056) (0.067) (0.063) Have you been affected by the pandemic? (negatively) 0.064* 0.184*** 0.083 -0.033 (0.033) (0.051) (0.050) (0.042) Observations 439 439 439 439 R-squared 0.083 0.215 0.206 0.176
Panel D
(1) (2) (3)
VARIABLES Did you take action
before the government?
Did your revenue decline in the
last two months?
How long before you expect to recover to
the pre-pandemic revenue?
Tobin’s q -0.009 0.003 -0.023* (0.025) (0.003) (0.013) Log of Total Assets 0.026 0.005 0.196*** (0.019) (0.018) (0.056) Cash / Total Assets -0.003 0.001 0.002 (0.003) (0.004) (0.013) Capex / Total Assets 0.014** 0.003 0.013 (0.006) (0.004) (0.014) Leverage 0.001 -0.000 0.006 (0.002) (0.002) (0.005) Sales Growth 0.020 -0.022 -0.082 (0.079) (0.055) (0.442) Have you been affected negatively by the pandemic? -0.092 0.142* 0.140 (0.063) (0.081) (0.211) Observations 267 208 137 R-squared 0.107 0.130 0.283 Country FE Yes Yes Yes Sector FE Yes Yes Yes
Notes: Robust standard errors are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1
52C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
expect a longer time until recovery. Firms with a higher Tobin’s q and thus higher
market/book valuation are more likely to reduce investment and more likely to need to
raise capital to weather the pandemic; further, they are less likely to voluntarily continue to
pay employees or services providers for disrupted services and make donations or shift
operations to fill pandemic needs, but expect a quicker recovery. Firms with higher sales
growth over the past 12 months are less likely to receive government support and more
likely to continue to pay employees or services providers for disrupted services. Firms with
higher cash reserves (relative to their assets) are less likely to need to raise capital to
weather the pandemic, while more leveraged firms are more likely to breach covenants on
their loans and reduce executive compensation but less likely to continue to voluntarily pay
employees or services providers for disrupted services. Firms with higher investment
spending before the pandemic are more likely to have adjusted their operation before
government guidance was issued.
In summary, financially more vulnerable firms see themselves more affected and reacted
more strongly to the pandemic, while larger firms acted more pro-actively to the threat of
the pandemic. We also find that firms more negatively affected by the pandemic are more
likely to have reduced investment spending, executive compensation and benefits, but are
also more likely to have continued to pay employees or services providers for disrupted
service; however, this might not indicate any causality but rather common shocks.
We run several additional analyses using the same regression set-up. First, for a subset of
firms we have information on the share of domestic and international revenues. While we
expect a higher negative impact and a faster reaction of more export-oriented firms, we do
not find any evidence for that, with the exception that firms with a higher share of domestic
revenue are more likely to state that they need to raise capital to weather pandemic and to
make negative changes to benefits, suggesting that the international trade channel was not
the primary channel through which the pandemic negatively impacted growth in emerging
markets; importantly, however, we regard these findings as tentative, given that we have
this variable available only for half of our sample.6 Second, we explore whether differences
in the pandemic’s impact and reaction of firms can be explained by variation in countries’
6 Results are available on request.
53C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
decision whether or not and when to impose lockdowns. Specifically, we consider the
difference between the date the firm returned the April survey and the date when the
government in the respective country imposed a stringent lockdown, which included school
closing, closure of non-essential businesses and mobility constraints. Only five of the ten
countries imposed such a strict lock-down, though at different dates, starting with the
Philippines on 17 March, followed by Malaysia on 22 March, South Africa on 26 March,
Vietnam on 1 April and Saudi Arabia on 6 April. The other five countries imposed less
stringent lockdowns between 22 March in Turkey and 10 April in Indonesia.7 We code the
length of time in lockdown as the date we received the survey response minus the date the
country entered strict lockdown if the country had imposed a strict lockdown by the time
we received the survey, and assign a zero to firms in countries which had not yet imposed a
strict lockdown as of the date we received their response. As most of the variation in these
regressions is explained at the country-level (with within-country variation determined by
the day of survey return), we drop country fixed effects from these regressions.
The results in Table 3 show that firms in countries that implemented a lock-down earlier are
more likely to be negatively affected by the pandemic, less likely to make negative changes
to benefits, and more likely to continue to voluntarily pay employees or services providers
for disrupted services and to provide financial flexibility to customers or business partners.
5. The delayed effect of firm responses on stock prices
We next gauge the effect of firms’ reactions to the crisis on firms’ stock prices. Specifically,
we regress the buy-and-hold stock return between 24 February and 24 April and between
24 April and 24 June on a set of country and sectoral fixed effects, the firm-characteristics
discussed above and the different survey questions from the early April round.8 Doing so,
7 The data on containment policies are from Olivier Lejeune’s web-site: https://github.com/OlivierLej/Coronavirus_CounterMeasures. We denote with strict lock-down any containment policy that has a five in the database (“All-day lockdown. Government requires citizens to shelter in place all day long. Citizens are allowed to come out to buy essential items.”) 8 We choose 24 April, as by this date, 98% of all first-round surveys had been received. 24 February is also the first trading date after a first local lockdown in Italy, thus can be seen as the date when the pandemic clearly had arrived in Europe. We also performed the regressions on a range of other dates to ensure the results were
54C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 3: COVID-19 Corporate Impact and Responses and Lockdown Dates
(1) (2) (3) (4) (5)
VARIABLES
Have you reduced
investment spending?
Made any changes to employee benefits? (negative)
Continued to pay
employees or services providers
for disrupted services?
Provided financial flexibility
to any customers or business partners?
Have you been
affected by the
pandemic? (negatively)
Difference of days between receiving the survey and the lockdown being implement 0.001 -0.003** 0.009*** 0.007*** 0.006*** (0.002) (0.001) (0.002) (0.002) (0.002) Have you been affected by the pandemic? (negatively) 0.133** 0.053** 0.186*** 0.076 (0.054) (0.022) (0.051) (0.050) Tobin’s q 0.014*** -0.007** -0.008** 0.003 -0.008
(0.003) (0.003) (0.003) (0.005) (0.006) Log of Total Assets 0.057*** 0.002 0.022 0.051*** 0.002
(0.014) (0.007) (0.014) (0.014) (0.013) Cash / Total Assets -0.003 -0.001 -0.001 -0.006** -0.000
(0.003) (0.001) (0.003) (0.002) (0.003) Capex / Total Assets 0.000 -0.001 0.001 -0.003 0.005
(0.005) (0.003) (0.006) (0.005) (0.005) Leverage 0.002 0.000 -0.004*** 0.001 0.001
(0.001) (0.001) (0.001) (0.001) (0.001) Sales Growth 0.113** 0.006 0.129** 0.057 0.085*
(0.056) (0.029) (0.058) (0.068) (0.048) Observations 439 416 439 439 439 R-squared 0.117 0.059 0.152 0.164 0.114 Country FE No No No No No Sector FE Yes Yes Yes Yes Yes
Notes: Robust standard errors are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1
we would like to assess how investors reacted to the differential impact of the crisis on
different firms and to their varied reactions. We also include each firm’s market beta, which
we estimate by regressing one year of weekly excess returns (from December 31, 2018
through December 31, 2019) on a constant and the weekly market factor from the
respective country.9
not sensitive to the specific date range and found consistent results. We do not include the questions from the June survey round. 9 We do not report the coefficient estimates for firm characteristics and market beta to save space, but results are available on request.
55C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Gauging stock market reaction between late February and late April (the initial outbreak
period of the global pandemic) allows us to test whether markets incorporated early
indications of the crisis impact and firm reactions, while gauging stock market reaction
between late April and late June (two months after 98% of survey responses had arrived)
allows to test the timely market reaction to public information.
The results in Table 4 (buy-and hold return between 24 February and 24 April) show that
markets priced in the need for firms to raise capital, reduction in investment, changes to the
dividend or share buyback, and layoffs. Specifically, firms that stated in early April that they
were reducing investment saw a 3.7% lower return, firms that have made changes to the
dividend or share buyback plan saw a 7.8% lower return, firms that needed to raise capital
to weather the pandemic saw a 5.8% lower return and firms that made layoffs saw a 11.1%
lower return. However, we do not find a significant market reaction to a negative or positive
impact of the pandemic.
Given that this information was available in early April, the efficient market hypothesis
would predict no significant market reaction to the impact of the pandemic or firms’
reactions. However, the results in Table 5 (buy-and hold return between 24 April and 24
June) show that firms that stated in early April that they were negatively affected by the
pandemic saw a 6.1% lower return than firms that were not or were even positively
affected by the pandemic, suggesting (together with the statistically insignificant results in
Table 4) that stock markets were slow to react to the negative impact of the pandemic on
firms. Firms which reported a positive impact averaged a return 18.3% higher the following
two months though this coefficient estimate is not significant, partially due to the rather
small sample of positively affected firms (only 13 such firms). Firms that stated in early April
that they were reducing investment saw a 6.1% lower return than firms that did not reduce
investment. Similarly, we find that firms that voluntarily provide financial flexibility to any
customers have 4.4% lower cumulative returns. This suggests that the market had a delayed
56C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 4: Stock Returns and COVID-19 Corporate Responses, February - April Panel A
(1) (2) (3) (4) (5) VARIABLES Cumulative Returns
Have you been affected by the pandemic? (positively) 0.102 (0.062)
Have you been affected by the pandemic? (negatively) -0.015 (0.020)
Have you or might you breach covenants? -0.071 (0.050)
Would you need to raise capital to weather the pandemic? -0.058** (0.023)
Do you expect to receive government support? 0.000 (0.019)
Observations 408 408 408 408 408 R-squared 0.144 0.142 0.127 0.127 0.133
Panel B (1) (2) (3) (4) (5)
Have you reduced investment spending? -0.037** (0.018)
Have you made changes to the dividend or share buyback plan? -0.078*** (0.023)
Have you made changes to executive compensation plans? -0.035 (0.024)
Have you made any layoffs? -0.111*** (0.028)
Made any changes to employee benefits? (negative) -0.005 (0.033)
Observations 408 408 408 408 386 R-squared 0.144 0.127 0.130 0.127 0.142
Panel C (1) (2) (3) (4)
Have you taken any measures to protect your stakeholders? 0.077*** (0.027)
Continued to pay employees for disrupted services -0.010 (0.019) Provided financial flexibility to customers? -0.006 (0.018) Made any donations to help fight the pandemic? 0.068***
(0.021)
Observations 408 408 408 408 R-squared 0.127 0.127 0.135 0.128 Country FE Yes Yes Yes Sector FE Yes Yes Yes Firm Controls Yes Yes Yes
Notes: The dependent variables are cumulative returns measured February 24th – April 24th. All regressions also include market beta as an additional firm control. Robust standard errors are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1
57C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 5: Stock Returns and COVID-19 Corporate Responses, April - June Panel A
(1) (2) (3) (4) (5) VARIABLES Cumulative Returns
Have you been affected by the pandemic? (positively) 0.183 (0.125)
Have you been affected by the pandemic? (negatively) -0.061** (0.026)
Have you or might you breach covenants? 0.008 (0.047)
Would you need to raise capital to weather the pandemic? -0.016 (0.027)
Do you expect to receive government support? -0.038* (0.022)
Observations 408 408 408 408 408 R-squared 0.144 0.142 0.127 0.127 0.133
Panel B (1) (2) (3) (4) (5)
Have you reduced investment spending? -0.061*** (0.023)
Have you made changes to the dividend or share buyback plan? -0.008 (0.026)
Have you made changes to executive compensation plans? -0.031 (0.026)
Have you made any layoffs? 0.004 (0.036)
Made any negative changes to employee benefits? -0.072** (0.032)
Observations 408 408 408 408 386 R-squared 0.144 0.127 0.130 0.127 0.142
Panel C (1) (2) (3) (4)
Have you taken any measures to protect your stakeholders? 0.017 (0.036)
Continued to pay employees or services providers -0.010 for disrupted services? (0.022) Provided financial flexibility to any customers -0.044** or business partners? (0.020) Made any donations to help fight the pandemic? -0.017
(0.025)
Observations 408 408 408 408 R-squared 0.127 0.127 0.135 0.128 Notes: The dependent variables are cumulative returns measured April 24th – June 24th. All regressions
also include market beta as an additional firm control. Robust standard errors are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1
58C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
reaction to the available information, not consistent with the efficient market hypothesis.
The results in Tables 4 and 5 also show a positive market reaction to a stakeholder-centric
approach by firms. Between late February and late April, firms that provided donations and
took voluntary measures to protect stakeholders had, on average, 6.8% and 7.7% higher
returns, respectively. Furthermore, firms that reduced employee benefits in April
subsequently saw 7.2% lower stock returns than other firms. We also do not find any
negative and significant reaction between late April and late June to firms providing
donation or voluntarily taking measures to protect stakeholders. This suggests that the
financial markets valued these stakeholder-centric corporations more than their
counterparts during the crisis and that the market has indeed priced in attempts of firms to
maintain long-term relationships of firms with employees, business partners and other
stakeholders.
6. Conclusions
This paper presented firm-level evidence on the effect of the COVID-19 pandemic in 10
emerging markets as well as firms’ reactions. Survey evidence shows an overall negative
effect, but also a quick reaction by firms. Adjustments are mostly on the investment rather
than labour side, with significant expectations of government support. We find quite some
variation in impact and firm reaction across not only countries and sectors, but also firms
with different characteristics.
Our results point to quick reaction of firms in emerging markets (often even before
governments), their focus on accommodating their labour force and maintaining long-term
relationships with stakeholders. We also find evidence of markets being slow to incorporate
publicly available information and valuing stakeholder centric activities. Our findings
provide a snapshot of the current situation and another stock-taking later this year might
provide different findings, but they show the picture of corporations trying to make the best
of a very bad situation.
59C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
References
Accetturo, Antonio and Anna Giunta (2018): Value Chains and the Great Recession: Evidence from Italian and German firms, International Economics 153, 55-68. Alfaro, Laura , Gonzalo Asis, Anusha Chari and Ugo Panizza (2019): Corporate debt, firm size and financial fragility in emerging markets, Journal of International Economics 118, 1-19. Alfaro, Laura, Anusha Chari, Andrew Greenland and Peter K. Schott (2020): Aggregate and firm-level stock returns during pandemics, in real time, Covid Economics, Vetted and Real-Time Papers 4, April. Alon, Titan, Minki Kim, David Lagakos , Mitchell VanVuren (2020): Lockdowns in developing countries should focus on shielding the elderly, VoxEU 25 June Bennedsen, Morten, Birthe Larsen, Ian Schmutte and Daniela Scur (2020): Preserving job matches during the COVID-19 pandemic: Firm-level evidence on the role of government aid, Covid Economics, Vetted and Real-Time Papers 27, June Bosio, Erica, Simeon Djankov, Filip Jolevski and Rita Ramalho (2020a), Survival of Firms during Economic Crisis”, World Bank Policy Research Working Paper No. 9239, Washington, DC. Bosio, Erica, Filip Jolevski, Joseph Lemoine and Rita Ramalho (2020b) Survival of firms in developing economies during economic crisis, in: Djankov, Simeon and Ugo Panizza (Eds): Covid-19 in Developing Economies. CEPR Press, London, UK Capelle-Blancard, Gunther and Adrien Desroziers (2020): The stock market is not the economy? Insights from the Covid-19 crisis, Covid Economics, Vetted and Real-Time Papers 28, June Ding, Wenzhi, Ross Levine, Chen Lin, and Wensi Xie (2020): Corporate Immunity to the COVID-19 Pandemic, NBER Working Paper 27055. Djankov, Simeon and Ugo Panizza, Eds. (2020): Covid-19 in Developing Economies. CEPR Press, London, UK. Gabaix, Xavier, 2011. The granular origins of aggregate fluctuations. Econometrica 79, 733–772. Gómez, Margarita, Andriy Ivchenko, Elena Reutskaja and Pablo Soto-Mota (2020): Behaviours, perceptions and mental wellbeing in high-income and low/middle-income countries at the beginning of COVID-19 pandemic, in: Djankov, Simeon and Ugo Panizza (Eds): Covid-19 in Developing Economies. CEPR Press, London, UK
60C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Hevia, Constantino and Andy Neumeyer (2020): A perfect storm: COVID-19 in emerging economies, in: Djankov, Simeon and Ugo Panizza (Eds): Covid-19 in Developing Economies. CEPR Press, London, UK IMF (2020): World Economic Outlook Update, June 2020: A Crisis Like No Other, An Uncertain Recovery Washington, DC Li, Kai, Xing Liu, Feng Mai and Tengfei Zhang (2020): The Role of Corporate Culture in Bad Times: Evidence from the COVID-19 Pandemic, Covid Economics, Vetted and Real-Time Papers 32, June. Rahman, H Z, N Das, I Matin, M A Wazed, S Ahmed, J Nusrat and U Zillur (2020), Livelihoods, Coping and Support during the COVID-19 Crisis, BGID-PPRC Rapid Response Research. Ramelli, Stefano and Alexander F. Wagner (2020): Feverish Stock Price Reactions to COVID-19. CEPR Discussion Paper 14511.
61C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Appendix A1: Survey questions
Early April survey
1. Have your operations been affected by the coronavirus pandemic; or do you expect them to be affected?
2. Have you breached, or might you breach, any covenants as a result? 3. Will raising additional capital likely be necessary to weather the crisis? 4. In response to the coronavirus pandemic, have you:
a. reduced or delayed any investment spending plans? b. made any layoffs? c. made changes to any employee benefits (e.g. health care, sick leave,
compensation, etc)? d. made any changes to executive compensation plans? e. made any changes to dividend or share buyback plans? f. taken any measures to protect employees, customers, or other stakeholders; or to
prevent the spread of the virus, in excess of government guidelines? g. voluntarily continued to pay any employees or services providers for services
which are temporarily disrupted? h. provided extraordinary non-contractual financial flexibility to any customers,
suppliers, or business partners? i. made any additional coronavirus-related donations (cash, supplies, human
resources), or shifted any business operations to fulfil crisis needs (e.g. manufacture of masks, hand sanitizer, etc) to help fight the pandemic?
j. received any coronavirus-related government support, or do you expect to be a beneficiary of any coronavirus-related government support?
Early June survey 1. In response to the coronavirus pandemic, have you: 2. Approx how much have revenues declined over the past two months? 3. By when do you expect a recovery to pre-crisis revenues? 4. Are there any updates to the answers you sent us in April (in the email below)?
62C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A2: Survey answers
April survey June update
Yes No No answer Yes New Operations affected 82% 15% 3% 82% 0% Breach covenants (or likely to breach) 1% 93% 6% 1% 0% Raising additional capital likely necessary 12% 80% 8% 28% 16% Received government support (or expect to receive) 35% 57% 8% 36% 1% Reduced or delayed investment spending plans 50% 41% 10% 50% 0% Layoffs 3% 91% 5% 7% 3% Changes to employee benefits 24% 70% 6% 27% 3% Changes to executive comp 17% 76% 7% 23% 5% Changes to dividend or share buyback plans 13% 80% 8% 14% 1% Measures to protect employees or other stakeholders 91% 5% 5% 91% 0%
Continued to pay employees for disrupted services 47% 45% 8% 50% 3% Financial flexibility to customers 41% 51% 9% 42% 2% Donations or shifted business operations 77% 18% 5% 77% 0%
Positive Negative Not clear Positive New
Operations affected 3% 69% 11% 3% 0% Changes to employee benefits 12% 7% 5% 15% 3%
Approx how much have revenues declined over the past two months? Total Decline 49% Increase 9% Minimal or no impact/change 13% No answer / declined to answer 28%
By when do you expect a recovery to pre-crisis revenues? Total 2020 24% 2021 12% 2022 or beyond 3% Maintaned or better performance 11% Not sure / unclear / no answer 50%
Did you take action before gov't pandemic measures were put in place, or following the issue of gov't measures? Total Before 63% Following 27% No clear answer 10%
63C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A3: Two-sample Tests for Equal Means
(a) Two-sample test for equal means:Firms that responded to April survey compared to those that did not
Obs 1 Obs 2 Mean 1 Mean 2 Dif SE t-value p-value Tobin’s q 99 439 1.179 1.351 -.172 .361 -.5 .634 Total Assets 99 439 6019.028 5700.712 318.317 1702.793 .2 .852 Cash / Total Assets 99 439 7.955 8.309 -.353 1.019 -.35 .729 Capex / Total Assets 95 439 4.806 4.489 .318 .579 .55 .584 Leverage 99 439 22.55 25.734 -3.184 2.126 -1.5 .135 Sales Growth 97 439 .197 .102 .095 .044 2.15 .032
Notes: Obs 1 is equal to one if companies did not respond to the April survey. Obs 2 is equal to 1 if companies responded to the April survey.
(b) Two-sample test for equal means:Firms that responded to April and June survey compared to those that only responded to April survey
Obs 1 Obs 2 Mean 1 Mean 2 Dif SE t-value p-value Tobin’s q 52 387 .956 1.405 -.449 .516 -.85 .385 Total Assets 52 387 3301.613 6023.071 -2721.45 2202.827 -1.25 .218 Cash / Total Assets 52 387 6.687 8.526 -1.839 1.315 -1.4 .163 Capex / Total Assets 52 387 5.795 4.313 1.481 .716 2.05 .039 Leverage 52 387 26.619 25.615 1.004 2.774 .35 .718 Sales Growth 52 387 .12 .1 .02 .05 .4 .69
Notes: Obs 1 is equal to one if companies responded only to the April survey. Obs 2 is equal to 1 if companies responded to both April and June survey.
64C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A4: Country variation in survey responses
Bangladesh Indonesia Malaysia Pakistan Philippines Saudi Arabia South Africa Thailand Turkey Vietnam Total June
updated Response rate 53% 73% 90% 56% 84% 67% 79% 91% 83% 80% 77% 85% Sample size 20 55 69 25 46 49 72 82 29 41 488 415 Operations affected 90% 75% 87% 72% 98% 73% 89% 79% 72% 78% 82% 82%
Positively affected 0% 0% 1% 4% 4% 2% 7% 1% 3% 2% 7% 3% Negatively affected 85% 55% 80% 68% 87% 63% 61% 68% 69% 61% 69% 69%
Breach covenants (or likely to breach) 0% 0% 0% 8% 0% 4% 3% 0% 0% 0% 1% 28% Raising additional capital likely necessary 45% 18% 13% 8% 11% 8% 1% 11% 14% 15% 12% 36% Received government support (or expect to receive) 40% 35% 39% 44% 22% 49% 19% 39% 34% 37% 35% 50% Reduced or delayed investment spending plans 55% 56% 45% 16% 61% 20% 63% 65% 34% 49% 50% 7% Layoffs 0% 5% 0% 0% 7% 6% 3% 4% 3% 5% 3% 27% Changes to employee benefits 10% 24% 17% 16% 37% 10% 25% 40% 7% 29% 24% 23%
Positive changes 5% 9% 12% 8% 20% 0% 10% 26% 3% 10% 12% 15% Negative changes 0% 5% 3% 8% 0% 4% 10% 16% 3% 12% 7% 7%
Changes to executive comp 10% 22% 12% 8% 11% 12% 25% 26% 14% 15% 17% 23% Changes to dividend or share buyback plans 10% 33% 7% 0% 13% 2% 17% 7% 24% 10% 13% 14% Measures to protect employees or other stakeholders 100% 95% 94% 100% 93% 88% 85% 93% 79% 83% 91% 91% Continued to pay employees for disrupted services 40% 42% 48% 60% 74% 33% 71% 28% 31% 41% 47% 50% Financial flexibility to customers 15% 29% 38% 36% 70% 20% 51% 48% 28% 44% 41% 42% Donations or shifted any business operations 80% 87% 83% 76% 91% 47% 78% 83% 59% 76% 77% 77% JUNE SURVEY Response rate 60% 87% 87% 84% 93% 80% 79% 94% 83% 83% 85% Sample size 12 48 60 21 43 39 57 77 24 34 415 Approx. how much have revenues declined over the past two months? Decline 58% 56% 35% 57% 63% 38% 40% 52% 38% 38% 49% Increase 0% 6% 5% 5% 7% 10% 7% 9% 8% 21% 9% Minimal or no impact/change 0% 4% 20% 24% 2% 13% 4% 9% 25% 21% 13% No answer / declined to answer 42% 33% 40% 14% 28% 38% 49% 30% 29% 21% 28% By when do you expect a recovery to pre-crisis revenues? 2020 8% 33% 15% 33% 19% 21% 5% 27% 25% 26% 24% 2021 17% 10% 13% 5% 16% 13% 9% 14% 8% 9% 12% 2022 or beyond 0% 4% 3% 0% 5% 3% 2% 8% 0% 0% 3% Maintained or better performance 8% 8% 15% 14% 0% 13% 2% 8% 21% 24% 11% Not sure / unclear / no answer 50% 42% 52% 43% 49% 36% 68% 38% 29% 29% 50% Did you take action before gov't pandemic measures were put in place, or following the issue of gov't measures? Before 50% 60% 62% 62% 56% 49% 58% 71% 33% 53% 63% Following 8% 35% 33% 14% 30% 33% 26% 19% 17% 21% 27% No clear answer 8% 4% 2% 5% 2% 3% 2% 5% 29% 21% 10%
65C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A5: Sector variation in survey responses
Comms Cons Discr Cons Staples Energy Fin Health Care Ind IT Mat Real Estate Utilities Total
June updated
Response rate 82% 75% 87% 81% 71% 85% 84% 82% 76% 76% 61% 77% 85% Sample size 23 65 47 25 77 17 70 31 66 53 14 488 415 Operations affected 87% 92% 74% 80% 70% 76% 90% 84% 77% 96% 50% 82% 82%
Positively affected 17% 3% 11% 0% 1% 6% 0% 0% 0% 0% 0% 7% 3% Negatively affected 61% 83% 51% 68% 52% 65% 83% 74% 65% 85% 43% 69% 69%
Breach covenants (or likely to breach) 4% 2% 0% 0% 0% 0% 1% 3% 2% 2% 0% 1% 28% Raising additional capital likely necessary 13% 18% 9% 28% 5% 12% 20% 6% 6% 11% 7% 12% 36% Received government support (or expect to receive) 39% 54% 23% 52% 17% 35% 49% 39% 32% 30% 0% 35% 50% Reduced or delayed investment spending plans 35% 58% 45% 64% 35% 41% 54% 55% 47% 64% 43% 50% 7% Layoffs 0% 9% 4% 0% 1% 6% 7% 0% 2% 2% 0% 3% 27% Changes to employee benefits 39% 29% 19% 24% 22% 24% 27% 19% 17% 30% 14% 24% 23%
Positive changes 39% 8% 11% 16% 16% 12% 10% 3% 8% 11% 14% 12% 15% Negative changes 0% 14% 2% 4% 3% 6% 13% 10% 8% 8% 0% 7% 7%
Changes to executive comp 4% 31% 9% 8% 12% 12% 26% 23% 12% 25% 0% 17% 23% Changes to dividend or share buyback plans 13% 18% 2% 20% 14% 12% 20% 13% 6% 8% 7% 13% 14% Measures to protect employees or other stakeholders 96% 89% 85% 92% 88% 82% 91% 100% 88% 94% 100% 91% 91% Continued to pay employees for disrupted services 48% 46% 45% 36% 52% 29% 44% 32% 56% 60% 21% 47% 50% Financial flexibility to customers 48% 35% 36% 24% 62% 18% 37% 45% 23% 58% 29% 41% 42% Donations or shifted any business operations 96% 74% 77% 76% 79% 88% 70% 71% 76% 79% 93% 77% 77% JUNE SURVEY Response rate 91% 83% 85% 88% 84% 82% 90% 97% 74% 85% 86% 85% Sample size 21 54 40 22 65 14 63 30 49 45 12 415 Approx. how much have revenues declined over the past two months? Decline 43% 67% 43% 55% 29% 43% 46% 33% 57% 51% 42% 49% Increase 14% 2% 15% 5% 11% 14% 8% 7% 10% 2% 8% 9% Minimal or no impact/change 10% 2% 10% 14% 14% 14% 16% 17% 14% 2% 25% 13% No answer / declined to answer 33% 30% 33% 27% 46% 29% 30% 43% 18% 44% 25% 28% By when do you expect a recovery to pre-crisis revenues? 2020 14% 26% 25% 14% 14% 14% 21% 20% 31% 24% 17% 24% 2021 5% 11% 8% 18% 9% 0% 19% 13% 8% 18% 8% 12% 2022 or beyond 0% 7% 3% 0% 2% 0% 8% 0% 0% 7% 0% 3% Maintained or better performance 10% 4% 20% 5% 6% 36% 10% 23% 8% 2% 17% 11% Not sure / unclear / no answer 62% 41% 38% 55% 54% 50% 37% 37% 51% 38% 50% 50% Did you take action before gov't pandemic measures were put in place, or following the issue of gov't measures? Before 81% 50% 60% 45% 57% 64% 54% 67% 55% 60% 83% 63% Following 10% 30% 23% 32% 25% 21% 30% 20% 39% 24% 0% 27% No clear answer 10% 2% 3% 9% 11% 7% 6% 10% 4% 7% 0% 10%
66C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
37-
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A6: R-Squared Analyses
Panel A
(1) (2) (3) (4)
VARIABLES Have you reduced investment spending?
Observations 439 439 439 439
Adjusted R-squared 0.054 0.098 0.040 0.133
Firm Controls Yes No No Yes
Country FE No Yes No Yes
Sector FE No No Yes Yes
Panel B
(5) (6) (7) (8)
VARIABLES Made any donations to help fight the pandemic?
Observations 439 439 439 439
Adjusted R-squared 0.056 0.067 0.003 0.124
Firm Controls Yes No No Yes
Country FE No Yes No Yes
Sector FE No No Yes Yes
Panel C
(9) (10) (11) (12)
VARIABLES Do you expect to receive government support?
Observations 439 439 439 439
Adjusted R-squared 0.026 0.018 0.051 0.062
Firm Controls Yes No No Yes
Country FE No Yes No Yes
Sector FE No No Yes Yes
Robust standard errors are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1
67
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Covid Economics Issue 38, 16 July 2020
Copyright: Harry Mamaysky
Financial markets and news about the coronavirus1
Harry Mamaysky2
Date submitted: 8 July 2020; Date accepted: 13 July 2020
I examine how financial markets interact with news about the COVID-19 pandemic. A twelve topic model optimizes the trade-off between number of topics and topic coherence. Using this model, I show that before mid-March 2020 markets react more to the same quantum of news when volatility is higher – a phenomenon I call hypersensitivity. Formal tests identify a structural break in mid-March, post which markets are no longer hypersensitive. In the hypersensitive stage, markets are overly volatile and overreact to news. Despite hypersensitivity, lagged prices better forecast future COVID-19 case counts than do lagged news.
1 I thank seminar participants from Columbia, QWAFAxNEW, UBS, and the Wolfe Research NLP and ML Conference, as well as Charles Calomiris and Paul Glasserman, for helpful suggestions.
2 Columbia Business School.
68C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
1 Introduction
In times of great economic and market stress, asset prices often respond dramatically
to news about underlying economic or market conditions. This was true during the
global �nancial crisis (GFC) of 2008�2009, when news about each successive multi-billion
dollar loss would cause �nancial markets to wildly gyrate. Yet during the GFC, news
data sets and the tools to analyze them were not yet widely available. Since the GFC,
the use of natural language processing (NLP) in �nance and economics has experienced
sizable growth. The COVID-19 pandemic of 2020 has emerged as the �rst global crisis
that can be analyzed in real-time using NLP tools developed in the last decade. Rather
than draw anecdotal conclusions about how news and markets interact, researchers can
systematically analyze the news-markets relationship. In this paper, I attempt such an
analysis.
Over the time period from January to April of 2020, I analyze the 72,263 Reuters
news articles that mention �coronavirus� or �COVID-19.� I �nd there are twelve distinct
topics into which these stories can be classi�ed. The choice of twelve topics optimally
balances the desire to have more topics to better represent news �ow with the requirement
that the topics are coherent � that is that these topics contain words that �sensibly�
belong together, as in Newman et al. (2010). The prevalence of certain topics in news
articles e�ectively tracks the narrative about the evolution of the pandemic. From early
discussions of the health and market impact of the coronavirus, the narrative shifts to a
discussion of the impact of the pandemic on corporations (2 topics), of its e�ect on credit
markets (2 topics), and of the policy responses by central banks and governments. The
other topics focus on currencies, European economies, oil and commodities, and sports.
As in Calomiris and Mamaysky (2019), I construct a measure of topical sentiment by
interacting topical frequency with sentiment, the latter measured using the Loughran and
McDonald (2011) dictionary. Topical sentiment is contemporaneously related to price
changes in the SP500 stock index, in the VIX volatility index (which measures implied
volatility on short-term SP500 options), in a high-yield corporate bond index (HY), and
in yield changes in 2- and 10-year US Treasuries. Daily topical sentiment is able to explain
a large portion of the variation in daily returns of these �ve asset classes. I argue that
a contemporaneous relationship between asset returns and news is to be expected under
the log-linear return approximation of Campbell and Shiller (1988) and Campbell (1991).
I check to see whether asset prices or the news-based measures are better forecasters
of future COVID-19 case counts. Given investor focus on understanding potential coro-
navirus outcomes over my sample period, contemporaneous market responses to news
69C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
would be justi�ed if news were a good forecaster of future case incidence. I �nd that
�ve of the top seven forecasters of future COVID-19 case counts are prices of the asset
classes mentioned above, which optimally forecast future case counts at lags of two to four
weeks. Lags of various news measures also forecast future case counts, but not as well as
the market price series. In all cases, lags are chosen to maximize the absolute value of the
correlation of the lagged series with future COVID-19 case counts. Lagged market prices
are generally better forecasters of future COVID-19 case incidence than lagged news.
The contemporaneous relationship between news measures and asset returns becomes
considerably larger when the VIX index has been elevated over the prior 10 trading
days. The same piece of news has a considerably larger e�ect on asset prices when lagged
implied volatility is higher. I refer to this phenomenon as hypersensitivity. Thus a portion
of the increased volatility during the coronavirus crisis is caused not by progressively
more extreme news �ow, but by progressively larger market responses to similar news
�ow. In the hypersensitive phase, market prices respond to dimensions of news �ow that
are ultimately uninformative about future fundamentals, as measured by incidence of
COVID-19 cases. At the same time, markets do not respond to dimensions of news that
prove to be informative about future COVID-19 cases. This raises the possibility that a
sizable amount of asset price movements during the hypersensitive phase were, in a sense,
unwarranted.
Aspects of this dynamic are suggested by the theoretical analysis in Glasserman, Ma-
maysky, and Shen (2020), who argue that extreme media scrutiny can push markets into a
high-volatility, low-price regime, which in turn invites more media scrutiny. In the model,
such regimes are characterized by prices that are extremely sensitive to fundamentals.
Part of the market's response to the global coronavirus pandemic appears to stem from
this kind of news-induced extreme sensitivity.
Structural break tests, using the Andrews (1993, 2003) distribution for the maximal
Chow statistic, suggest that a break in the contemporaneous news-markets relationship
occurred around the middle of March 2020. This was a time of extreme market stress,
as well as of extreme responses by central banks and �scal authorities. I speculate that
such policy actions contributed to the structural break in the news-markets relationship,
suggesting an additional role for monetary and �scal policy in averting adverse market
outcomes: namely, non-virtuous news-market cycles can potentially be broken by ap-
propriate policy responses. The hypersensitivity dynamic is much more prevalent in the
pre-break subsample. After the structural break occurs, there is very little evidence of hy-
persensitivity. News have the same e�ect on contemporaneous asset price changes during
70C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
high- and low-VIX periods in the post-break regime.
I document the presence of lead-lag relationships between news and markets. Today's
market moves Granger cause tomorrow's topical sentiment. Generally today's news �ow
is worse when yesterday's market returns were poor, even after controlling for yester-
day's news �ow. This suggests that news coverage, even of topics that are not directly
market-related, is tainted by past poor market performance. Interestingly, today's topical
sentiment also Granger causes tomorrow's market moves. This is true even after control-
ling for today's and yesterday's market moves. The direction of news forecastability for
tomorrow's market moves overwhelmingly suggests overreaction, i.e. a price move tomor-
row which reverses a portion of today's price response to a piece of news. The Granger
causality network is considerably denser in the pre-break than in the post-break regime.
Furthermore most of the price overreaction occurs in the pre-break regime, and almost
all of the overreaction occurs for hypersensitive news-market pairs. Hypersensitive price
reactions apparently are quickly, at least partially, reversed. In the post-break subsam-
ple, there is much less overreaction, and none of it happens in response to hypersensitive
news-market pairs.
The substantially di�erent contemporaneous and lead-lag relationships that prevail
in the pre- and post-break regimes suggest that the markets response to the coronavirus
crisis consisted of two distinct phases: an early hypersensitive phase, and a late relatively
calm phase. In the hypersensitive phase, markets hung on every mention of coronavirus
and COVID-19 as they tried to assess the economic impact of the pandemic. In the post-
break phase, the sensitivity to coronavirus news �ow diminishes. The strong evidence
of overreaction in the early, hypersensitive phase suggests a portion of contemporaneous
market response to coronavirus news �ow was unwarranted.
The rest of the paper proceeds as follows. Section 2 describes the text analytics.
Section 3 describes the non-text data sets. Section 4 studies the contemporaneous rela-
tionship between news �ow and asset returns. Section 5 analyzes the lead-lag structure
between news and market returns. Section 6 o�ers some interpretations and discusses a
mechanism that may explain some of the �ndings. Section 7 concludes. Section A.2 of
the appendix discusses some robustness tests.
1.1 Connection to the literature
The paper most closely related to the present work is Baker et al. (2020), who �nd that
the majority of stock market moves greater than 2.5% from February 24 to March 24,
2020 were attributed by newspapers to news about COVID-19. I look more closely at the
71C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
dynamics of the news-markets relationship, and do not focus only on large price move
days. Furthermore, I analyze other markets (HY, Treasuries, VIX) besides the SP500
stock index. Loughran and McDonald (2020) document that in 2018 only 21% of �rms
mention pandemic-related risks in their 10-K �lings, and wonder why this number wasn't
higher. Gormsen and Koijen (2020) use equity dividend futures and the level of the stock
market to extract changes in investor expectations about economic growth in response to
the coronavirus pandemic. Altman (2020) investigates the impact of the COVID-19 crisis
on the credit cycle. My work is related to the two latter papers in that I document the
evolution of the narrative about the coronavirus pandemic, and show how the narrative
and markets interact with each other; this interaction can in turn impact how investors
think about future economic growth and corporate creditworthiness.
Roll's (1988) famous observation that monthly stock returns are di�cult to explain
even with contemporaneous in�uences precipitated much interest in the �nance profession
to understand which events can be associated with stock price movements. Fair (2002)
identi�es sixty-nine events that led to large SP500 moves between 1982 and 1999, and
�nds that 53 of these were associated with monetary policy. Tetlock, Saar-Tsechansky,
and Macskassy (2008) show that contemporaneous market returns are more sensitive to
news sentiment when the word �earn� is mentioned in the news story. I �nd that market-
wide reactions at the asset class level to contemporaneous news about coronavirus are
larger in high-VIX periods. Furthermore, the R2s in my regressions of market returns on
contemporaneous news measures are quite high, suggesting that in the early part of the
COVID-19 crisis markets movements were, to a meaningful extent, driven by news about
coronavirus. Boudoukh, Feldman, Kogan and Richardson (2018) regress daily squared
stock returns on an indicator variable for the presence of contemporaneous intraday and
overnight news, and conclude that the presence of news increases daily return variance.
My results suggest that asset-level return variance during high-VIX days is at least par-
tially caused by increased sensitivity of asset returns to news �ow, and not because of the
arrival of more extreme news.
The evidence on overreaction in this paper is in contrast to the �nding that news
sentiment can predict aggregate stock returns (Tetlock 2007), and especially during reces-
sions (Garcia 2013). Both Tetlock (2007) and Garcia (2013) �nd evidence of index-level
underreaction to news. My analysis di�ers from theirs because I consider multiple asset
classes, and not only the US stock index, and because I focus on one particular high-stress
market episode whereas they analyze a much longer time-series that is not solely focused
on market crises.
72C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
2 Text analysis
The news articles used in this study are obtained from the Thomson-Reuters news archive,
and include all English-language articles in 2019 and through the end of April 2020 that
mention the terms �coronavirus� or �COVID-19�. There is only one article that mentions
coronavirus all of 2019.1 The �rst article mentioning coronavirus in 2020 is on January
8, and the coverage begins in earnest starting on January 17, the date that I start my
analysis. Figure 1 shows the daily count of articles mentioning coronavirus rises steadily
from January 17, 2020 onwards. When there are multiple articles in a given Primary
News Access Code chain,2 the last one is selected. Day t articles must have timestamps
prior to or equal to 4pm NY time; day t articles with later timestamps are classi�ed as
day t+ 1 articles. Post-4pm Friday, as well as Saturday and Sunday articles, are counted
as occurring on the subsequent Monday. Articles occurring after 4pm NY time on April
30, 2020 are excluded from the analysis. The �nal corpus contains 72,263 articles.
Article j's sentiment is given by
Sentj =Posj −Negj
Totalj,
where Posj and Negj are the number of positive and negative Loughran-McDonald (2011)
words appearing in the article, and Totalj is the total number of words after excluding
stop words. I employ the Das and Chen (2007) algorithm to mark negated words,3 and
negated sentiment words are ignored in the above counts. Daily aggregate sentiment Sentt
is the average Sentj of all articles classi�ed as day t articles. This is the series labeled
sent in Figure 1. The average value of daily sentiment in the corpus is -0.019 (Table 1).
Glasserman, Li, and Mamaysky (2020) show that, in a corpus of approximately 1.4 million
articles about SP500 �rms from 1996 to 2018, the average article sentiment is -0.011. The
sentiment in the present corpus is lower, though not dramatically so. I also calculate the
daily standard deviation of article-level sentiment; this series is labeled sent_sd.
The topic model is obtained by running latent Dirichlet allocation (LDA) on the
document-term matrix obtained by �rst stemming all the words in a news article, then
dropping stop words and several other commonly occurring words,4 and then keeping
1It is on February 27, 2019 and describes the Coalition for Epidemic Preparedness Innovations pre-sciently partnering with a German �rm to create a faster process for anti-pandemic vaccine production.
2This is how Reuters identi�es a collection that includes the initial article, and subsequent revisions.3Implemented via the mark_negation function of Python's NLTK package.4These are: said, thomsonreut, https, tmsnrt, www, reuter, coronavirus, com, nl.
73C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Text series and coronavirus case counts
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.028
0.026
0.024
0.022
0.020
0.018
0.016
0.014
0.012 sent
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.014
0.016
0.018
0.020
0.022
0.024
sent_sd
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0
500
1000
1500
2000
count
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.02
0.04
0.06
0.08
0.10f_sports
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.02
0.04
0.06
0.08
0.10
0.12
f_central bank
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.000
0.025
0.050
0.075
0.100
0.125
0.150
0.175
0.200f_markets
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.2
0.4
0.6
0.8
f_health
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.02
0.04
0.06
0.08
0.10
0.12 f_europe
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.02
0.04
0.06
0.08
f_oil & comm
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.05
0.10
0.15
0.20f_currency
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.01
0.02
0.03
0.04
0.05 f_credit
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.02
0.04
0.06
0.08
0.10
0.12 f_corp & govt US
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.05
0.10
0.15
0.20
0.25
0.30
f_corp actual
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.025
0.050
0.075
0.100
0.125
0.150
0.175
0.200
0.225
f_corp future
01-15
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0.00
0.01
0.02
0.03
0.04
0.05 f_credit1
02-01
02-15
03-01
03-15
04-01
04-15
05-01
0
20
40
60
80
100 corona
Text and coronavirus data from 2020-01-17:2020-04-30
Fig. 1. Summary of sentiment (sent), article count (count), topic frequencies (markedf_), and coronavirus con�rmed cases. Data are daily. The text measures are from allarticles that mention �coronavirus� or �COVID-19.� The solid lines correspond to textmeasures obtained from all articles in the corpus, and the dashed lines correspond to thesame text measures derived after all non-weekend intraday (timestamp from 9:30am�4pmNY time) articles have been removed.
74C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 1
Summary statistics for the variables used in the analysis. All data are daily. The labelssent, sent_sd, and count refer respectively to the daily mean and standard deviation ofarticle sentiment, as well as the daily number of articles. The f_[topic] and s_[topic]series refer to topical frequency and sentiment respectively. The coronavirus case counts,labeled corona, are obtained from the Johns Hopkins Coronavirus Resource Center, andare reported in thousands. N shows the numbers of days for which there are observations.The AR(1) (�rst-order autocorrelation) coe�cient is shown by default for the level of eachseries, but if marked with r (d) it is shown for one-day returns (di�erences). The marketsseries are labeled as follows: sp500 is the SP500 stock index; vix is the VIX volatilityindex; gt2 and gt10 are the 2- and 10-year Treasury yields respectively; and hy refers tothe FTSE US high-yield index. The series vix_avg refers to the average level of vix overthe prior ten trading days.
Summary statistics for daily variables
Variable N Mean StdDev 5% 25% Median 75% 95% AR(1)
sent 75 -0.019 0.002 -0.023 -0.021 -0.019 -0.018 -0.015 0.268sent_sd 75 0.021 0.002 0.017 0.021 0.022 0.023 0.024 0.806count 75 960.840 553.533 203.100 439.500 993.000 1348.500 1793.000 0.888f_sports 75 0.040 0.019 0.008 0.027 0.042 0.050 0.070 0.841f_central bank 75 0.082 0.041 0.015 0.045 0.102 0.119 0.127 0.953f_markets 75 0.084 0.039 0.046 0.050 0.073 0.115 0.163 0.854f_health 75 0.223 0.108 0.143 0.170 0.187 0.252 0.312 0.644f_europe 75 0.065 0.014 0.051 0.061 0.065 0.070 0.087 0.361f_oil & comm 75 0.054 0.017 0.030 0.048 0.054 0.061 0.091 0.796f_currency 75 0.083 0.042 0.037 0.048 0.073 0.113 0.149 0.783f_credit 75 0.016 0.013 0.001 0.006 0.010 0.025 0.037 0.882f_corp & govt US 75 0.084 0.025 0.042 0.062 0.090 0.103 0.118 0.847f_corp actual 75 0.138 0.065 0.023 0.092 0.136 0.181 0.242 0.835f_corp future 75 0.109 0.047 0.053 0.069 0.098 0.155 0.188 0.825f_credit1 75 0.020 0.014 0.002 0.008 0.014 0.033 0.046 0.921s_sports 75 -0.001 0.000 -0.001 -0.001 -0.001 -0.000 -0.000 0.839s_central bank 75 -0.002 0.001 -0.003 -0.002 -0.002 -0.001 -0.000 0.938s_markets 75 -0.002 0.001 -0.003 -0.002 -0.001 -0.001 -0.001 0.767s_health 75 -0.004 0.002 -0.006 -0.005 -0.004 -0.003 -0.002 0.546s_europe 75 -0.001 0.000 -0.002 -0.001 -0.001 -0.001 -0.001 0.436s_oil & comm 75 -0.001 0.000 -0.001 -0.001 -0.001 -0.001 -0.001 0.643s_currency 75 -0.002 0.001 -0.003 -0.002 -0.001 -0.001 -0.001 0.689s_credit 75 -0.000 0.000 -0.001 -0.000 -0.000 -0.000 -0.000 0.863s_corp & govt US 75 -0.002 0.001 -0.002 -0.002 -0.002 -0.001 -0.001 0.783s_corp actual 75 -0.003 0.001 -0.005 -0.004 -0.003 -0.002 -0.000 0.840s_corp future 75 -0.002 0.001 -0.004 -0.003 -0.002 -0.001 -0.001 0.657s_credit1 75 -0.000 0.000 -0.001 -0.001 -0.000 -0.000 -0.000 0.914corona 70 32.092 34.827 0.528 2.216 9.303 71.505 83.783 0.980sp500 73 88.969 10.066 72.769 82.022 88.287 99.374 101.874 r -0.425vix 73 36.666 19.561 12.952 16.050 37.760 47.300 73.388 d -0.397hy 76 93.154 6.834 80.993 88.445 93.675 99.699 100.381 r 0.336gt2 76 0.784 0.548 0.204 0.241 0.530 1.414 1.530 d 0.014gt10 76 1.108 0.437 0.599 0.705 0.982 1.584 1.771 d -0.124vix_avg 76 34.846 19.238 12.734 15.654 37.593 49.560 67.167 d 0.859
75C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
all words occurring more than 10 times in any month's set of coronavirus articles.5 An
LDA topic is a probability distribution over, in this case stemmed, words. The interested
reader can consult Steyvers and Gri�ths (2007) for a good introduction to topic models.
Figure 2 shows word clouds that summarize the identi�ed topics, where each word's size is
drawn in proportion to its probability weight in a given topic. The topic labels re�ect my
qualitative judgment about the meaning of each topic. A twelve topic model is optimally
chosen, for reasons that I will discuss shortly. For example: the central bank topic contains
words such bank, govern, and billion; the credit topic contains words like credit, �tch (the
rating agency), and rate; there is an additional credit-related topic, which I label credit1.6
LDA assigns to each document a probability distribution over each of the topics in
the model. For example, a news article that discusses the credit impact of central bank
policy may have a 75% allocation to the central bank topic, a 20% allocation to the credit
topic, with the remaining 5% allocated to the remaining topics. To get a better sense
of the types of articles that fall into each topic, Table 5 in the appendix shows eight
headlines of articles that are representative of each topic: four each for the highest and
lowest sentiment articles among representative articles for each topic. An article is said
to be representative about topic i if its document-topic allocation to i is above 70%. In
the above example, the article would be representative of the central bank topic. As
can be seen from the table, most of the representative articles about the central bank
topic are appropriately classi�ed, as are all of the credit and credit1 articles. Most other
representative article headlines also appear appropriate. Finally, sentiment allocation of
the headlines appear intuitive: higher sentiment headlines generally convey more positive
news than negative sentiment ones.
5I use the LDA implementation in Python's scikit-learn package.6As discussed below, I tried many independent topic estimation runs, and many of these identi�ed
two distinct credit-related topics. Some runs identi�ed more than two.
76C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
central bank mean(coh)=0.302 med(coh)=0.282 corp actual mean(coh)=0.301 med(coh)=0.278
corp & govt US mean(coh)=0.185 med(coh)=0.138 corp future mean(coh)=0.351 med(coh)=0.333
credit mean(coh)=0.875 med(coh)=0.927 credit1 mean(coh)=0.921 med(coh)=0.932
77C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
currency mean(coh)=0.303 med(coh)=0.308 europe mean(coh)=0.215 med(coh)=0.195
health mean(coh)=0.351 med(coh)=0.342 markets mean(coh)=0.358 med(coh)=0.351
oil & comm mean(coh)=0.253 med(coh)=0.243 sports mean(coh)=0.295 med(coh)=0.287
Fig. 2. Word clouds describing topics associated with articles that mention �coronavirus�or �COVID-19.� Each word cloud is labeled with the topic name, as well as the mean andmedian coherence for the topic (as described in Section 2.1).
78C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 2
Averages of document-topic distributions (the �rst twelve rows), sentiment (labeled sent),and article counts (labeled Number), grouped by time period: intra refers to articlesthat come out from 9:30am�4pm on non-weekend days, and over refers to overnight andweekend articles. By construction, the �rst twelve entries in each column sum to one.
Average document-topic distribution by time-period
bucket intra over
sports 0.049 0.047
central bank 0.107 0.100
markets 0.053 0.075
health 0.201 0.203
europe 0.065 0.066
oil & comm 0.063 0.049
currency 0.068 0.066
credit 0.027 0.017
corp & govt US 0.117 0.085
corp actual 0.122 0.185
corp future 0.090 0.089
credit1 0.036 0.020
sent -0.020 -0.019
Number 21,640 50,623
For day t, I aggregate the document-topic distributions of all articles classi�ed as be-
longing to that day into a daily measure of topic frequency. Letting fj,k be the probability
allocation of article j to topic k, the daily topic frequency measure is
ft,k =1
Nt
∑j∈{day t articles}
fj,k, (1)
where Nt is the number of articles in day t. Table 2 shows average document-topic
distributions, sentiment, and article counts for intraday (9:30am�4pm, non-weekend) and
overnight periods (all others). The most prevalent topics are: health (top words include
case, report, virus); two corporate topics corp & govt US and corp actual (top words
include brief, company, report, buzz, corp); and central bank. The largest discrepancy
between within day topic incidence is in corp actual which is more likely to occur overnight
than intraday. These series are labeled f_[topic] in subsequent analysis. I discuss the
79C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
evolution of topic prevalence in Section 2.2.
Calomiris and Mamaysky (2019) �nd that topical sentiment is an important forecaster
of country-level returns. Topical sentiment is the product of topic frequency with senti-
ment at the daily level. This measure captures the extent to which negative or positive
news �ow is speci�c to a given topic. The daily topical sentiment series for topic k is
Sentt,k = Sentt × ft,k. (2)
For example, if day t has very negative daily sentiment Sentt, and has articles predomi-
nantly about central bank and credit topics, the topical sentiment of those two topics on
day t would be very negative, while the topic sentiment of the non-prevalent topics would
be close to zero. These series are labeled s_[topic] in subsequent analysis.
2.1 Topic model selection and coherence
It is well known that there can be multiple LDA topic models that result in either identical
or very similar log likelihoods for a given text corpus.7 Because of this indeterminacy care
must be taken to identify an appropriate topic model for a particular application. To select
an optimal model, I estimate 10 independent runs for eight di�erent choices of number
of topics: 3, 6, 9, 12, 15, 18, 21 and 24. This results in 80 topic models estimated using
independent runs of scikit-learn's LDA algorithm, each of which starts at a di�erent
random seed and ends at di�erent model estimates. Intuitively, the optimization here
involves choosing the maximum number of topics, which will be helpful in identifying
particular market-topic relationships, while ensuring that each of the topics is �sensible�
and not overly esoteric.
The natural language processing literature has identi�ed a useful measure of topic
quality, called coherence. Newman et al. (2010) show that various algorithmic measures of
topic coherence �t well with human evaluation of topic quality. Following their approach,
I use a coherence measure that utilizes cosine similarities among the top 10 most frequent
words in a given topic. Let D be the subset of the document-topic matrix for the present
corpus that is restricted to the top 1,000 most frequently occurring words.8 D is then a
7See Michal et al. (2004) for a convenient form of the LDA likelihood function for a text collection.Ke, Montiel Olea, and Nesbit (2020) give results on non-identi�cation of LDA topic models.
8This restriction only applies to the coherence calculation, and not to the LDA estimation which usesthe entire document topic matrix.
80C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
72,263 × 1,000 matrix. The cosine similarity of two words i and j is de�ned as
ci,j =D>i Dj‖Di‖‖Dj‖
,
where Di is the ith column of D. ci,j is a number from 0 to 1 and measures the tendency
of two words to co-occur. When ci,j = 1 then in every document in which i appears, so
does j, and in the same relative proportion; and if ci,j is zero, then i and j never appear
in the same document. If either i or j is not in the top 1,000 words (which happens very
infrequently in the analysis), then ci,j is set to zero.
The coherence of a given topic is then the mean or median of the(102
)= 45 possible
cosine similarities ci,j's among the top 10 words in a topic. High coherence topics consist
of top words that co-occur frequently in the corpus. Figure 2 shows the mean and me-
dian coherence of each of the twelve topics in the model. credit and credit1 are the two
highest coherence topics � the top words in each, �tch rate, credit � tend to co-occur very
frequently. On the other hand, corp & govt US is a low coherence topic. Its top words,
compani, trump, buzz, new, report, are less closely related and tend to co-occur much less
frequently than the credit and credit1 topics' top words. sports is a medium coherence
topic, whose top words sport, event, game, postpon, co-occur relatively frequently.
While the mean and median topic coherence were, in all cases I examined, very similar,
I chose the median as my base measure because it is less sensitive to outlier pairwise cosine
similarities. To rank a given topic model with K topics, I calculate the average median
coherence across the K topics. I refer to this as the coherence of a topic model, and write
it as CK . In the present corpus, there is evidence of a power law that prevails between
topic model coherence and number of topics. In a regression of log CK on logK (recall
there are 10 runs for each K), I �nd that
log CK = −0.3967− 0.2613 logK + noise,
where the White standard error of −0.2613 is 0.007 and the adjusted R2 of the regression
is 0.901. Given the high R2 of the �t, the quantity CKK0.25, or the scaled coherence of
the model, should equal a constant times a small error term. This quantity captures the
trade-o� between having a higher number of topics (desirable for �nding useful market-
topic relationships) and the observation that in the present corpus a higher number of
topics decreases average topic coherence (undesirable).
Figure 3 summarizes the results of this analysis. It plots the scaled topic coherence of
81C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
3 6 9 12 15 18 21 24Number of Topics K
0.60
0.65
0.70
0.75
0.80
K0.
25×
Cohe
renc
e
2nd best run:Topics 3=1Topics 6=4Topics 9=7Topics 12=6Topics 15=1Topics 18=4Topics 21=1Topics 24=8
Scaled coherence of LDA Models
run_num11023456789
Fig. 3. This �gure shows the scaled average of median coherence, as explained in Section2.1, for all topics across di�erent topic con�gurations and di�erent model runs. The x-axisshows the number of topics for a given set of model runs. The y-axis shows the scaledaverage coherence for a given topic model, as a function of the run number and numberof topics in the model. The table shows the run with the second best scaled averagecoherence for each topic con�guration.
82C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
all 10 runs for each K. As can be seen, except for model run 4 in the 9-topic set of models,
the highest scaled average coherence is obtained in the 12- and 15-topic runs. The 9-topic
run number 4 has a very high coherence because the model chooses three credit topics
instead of the usual two, and credit topics have high coherence as already mentioned. To
mitigate the e�ect of such outlier topic choices, I focus instead on the second best model
run across the Ks. The winning model run based on the second best criterion is run 6 in
the 12-topic model. This is the topic model that I use in the remaining analysis.
It should be noted that the main conclusion of the paper are not sensitive to the choice
of number of topics or to which run is chosen for a given topic number. All model run and
topic con�guration variants I have tried lead to the same qualitative conclusions. That
being said, more coherent topics increase the intepretability of the results. Ultimately,
the economic intuition behind the markets-news relationship depends on having topics
that are easy to understand and interpret. For this reason my model selection approach
should prove useful for future research. Users of topic models in economics should explore
the model space and �nd a suitable model for their particular application. The usual
approach of running a single LDA with a guess as to the appropriate topic number is
unlikely to yield the best-possible model.
2.2 Evolution of the narrative
Figure 1 shows the evolution of the daily topical frequencies ft,k in equation (1) from
January 17 through the end of April of 2020, the daily sentiment Sentt, the daily standard
deviation of article-level sentiment (sent_sd), as well daily article counts (count). The solid
lines in the �gure correspond to text measures derived from all articles in the corpus. The
orange, dashed lines correspond to text measures which exclude all intraday articles; these
will serve as a robustness test, as I discuss in Section A.2.
The aggregate sentiment series exhibits a lot of volatility in the early part of the
sample and achieves its minimum level around the middle of March of 2020. The overall
sentiment of all coronavirus articles improves from this low point in March through the
end of April. The daily standard deviation of article-level sentiment sent_sd starts very
low in January, peaks in the middle of March, and begins to slightly decline through April;
though it remains at very elevated levels relative to the start of the sample. The article
count starts at almost zero in early January, peaks at over 2,000 articles per day in the
middle of March, and then settles into the 1,000-1,500 range by the end of April.
Table 2 shows the average topic prevalence in the intraday and overnight periods, and
Figure 1 shows that topic incidence has varied meaningfully over the span of the crisis.
83C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
The news stories in the �rst few days of the crisis fall into the health topic, whose values
on January 17th and 18th were 87% and 69% respectively, before falling and staying in
the 20% range for the rest of the sample. On January 17th, there was virtually no news
coverage of markets, though this quickly changed by the next week, when 17% of all news
coverage regarding coronavirus falls into the markets topic. There was a spike in news
coverage about oil and commodities in early February, and this series then remains in
the mid-single digits for the rest of the period . Much of the early news coverage of the
pandemic dealt with its e�ect on currency markets, though the currency topic becomes
less prominent towards the end of the sample. The share of coverage of the European
impact from coronavirus remains steady at around the 6% level throughout the crisis.
The two corporate topics exhibit somewhat di�erent trends. The corp future topic,
which is largely about anticipation of future corporate impacts from the crisis, peaks in
the early part of the sample and then settles into the 5% to 10% range. The corp actual
topic, which deals with the realized e�ect of the pandemic on corporations, starts at a
very low level, and then climbs to be the most prevalent topic towards the middle and
end of the sample, as anticipated outcomes become realized. The least coherent topic,
corp & govt US, which seems to be about government and corporate interactions around
the coronavirus crisis, start at 5% and grows to roughly 10% of the news share. Given
the topic's low coherence, it is perhaps a catch-all topic for stories that do not neatly �t
into the other categories.
The sports topic starts out very low, as most early news coverage of the pandemic does
not deal with sports, then peaks at close to 10% of the news �ow in early March as all
major professional sports suspend their seasons, and as talk about postponing the Tokyo
Olympics intensi�es. The central bank topic is almost never mentioned in coronavirus
articles at the beginning of the sample, and then steadily grows until it represents roughly
12% of news �ow by early March, a level it maintains until the end of the sample. The
two credit topics are also absent in the early crisis news �ow and then grow steadily as
the profound impact of quarantine on corporate and consumer credit becomes apparent.
In the latter part of the sample the two credit topics, credit and credit1, represent just
under 10% of the news �ow.
In summary, the twelve topic model does a good job of tracking the evolution of news
�ow about the crisis, from the early articles dealing with the health impact of coronavirus
to the later articles dealing with central bank interventions and credit impacts. The hope,
therefore, is that these news measures will allow us to understand how markets and news
�ow about the coronavirus evolve throughout the crisis.
84C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 1 gives summary statistics for these text series, as well as for topical sentiment,
Sentt,k, which is the product of daily topic frequency and sentiment. The text series are all
fairly persistent as measured by their daily AR(1) coe�cients. The aggregate sentiment
series Sentt is less so.
3 Other data
I collect daily price data on the SP500 index, the VIX index, the FTSE US High-Yield
Market index (HY), which tracks the performance of high-yield corporate bonds, and US
2- and 10-year Treasury yields, labeled respectively gt2 and gt10. I include high-yield
bonds because they are a fairly liquid and particularly economically sensitive asset class.
The data are from Bloomberg, and go from January through the end of April 2020.
I obtain data on global con�rmed COVID-19 cases from the Johns Hopkins Coron-
avirus Resource Center. These data are updated daily, and start in 1/22/2020. According
to the GitHub page which stores the data,9 the update frequency is: �Once a day around
23:59 (UTC).�10 For example, the data for 4/1/2020 would be released at 23:59 UTC,
which is 7:59pm NY time. This is after the index calculation time on 4/1/2020, which
is 3pm for Treasuries and 4pm for SP500 and VIX, and perhaps an hour or two later
for high-yield.11 For these markets, the reaction to the 4/1/2020 COVID-19 case counts
would not happen until the next trading day's close, or 4/2/2020. For this reason I use
the day t increase in the global COVID-19 case count as the day t+ 1 value of my corona
series. This aligns the case counts with the days on which markets would have been able
to react. For Mondays, the one-day increase comes from Sunday's case counts. Assigning
to Mondays the cumulative case count increase from Friday, Saturday and Sunday would
introduce a day-of-week e�ect, which using only the Sunday increase avoids.
Figure 1 shows the corona series, reported in thousands of cases. The daily COVID-19
case counts start relatively low, and experience exponential growth starting in mid-March.
The spike in the series in February corresponds to a one-time restatement of case counts in
China. Table 1 gives summary statistics for the markets and corona series. For the table,
the HY and SP500 indexes are normalized to a value of 100 on January 16, 2020. The
corona series, though it represents the daily increase in case counts, is highly persistent
with an AR(1) coe�cient of 0.980. The markets series, whose AR(1) coe�cients measure
9I use the �le time_series_covid19_confirmed_global.csv fromgithub.com/CSSEGISandData/COVID-19/tree/master/csse_covid_19_data/csse_covid_19_time_series.
10UTC stands for Coordinated Universal Time, the successor standard to Greenwich Mean Time.11According to the FTSE Fixed Income Index Guide most indexes use closing prices from 3pm�6pm.
85C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
the autocorrelation of either returns (labeled with r) or di�erences (labeled with d), are
also unusually persistent. The AR(1) coe�cient for daily SP500 returns, for example, is
-0.425, suggesting a very large degree of mean-reversion over the sample. For context,
the AR(1) coe�cient for daily SP500 returns in 2019 was -0.09. The VIX index is also
strongly mean-reverting over this sample, while HY index returns are strongly positively
auto-correlated. Interestingly Treasury yield changes have very low autocorrelations.
3.1 A �rst look at the data
News �ow (to an extent) and �nancial markets are both forward looking. From Figure
1, it is clear that the media began to discuss the impacts of coronavirus long before the
COVID-19 case count began its exponential growth. This leads to the question of how
well and with that lead time did news �ow and markets anticipate future case incidence?
Figure 4 investigates this question. The text series used in the subsequent analysis
are the daily topical sentiment series Sentt,k from (2). I use topical sentiment instead
of topic frequency following Calomiris and Mamaysky (2019), who �nd topical sentiment
useful for forecasting country-level stock returns. Topical sentiment may be particularly
relevant because it captures both the prevalence of news �ows about a particular topic,
as well as the sentiment of that news �ow.
For each text and market series, I �nd the lag, between 0 and 21 trading days, at
which that series is most correlated (in absolute terms) with daily COVID-19 case counts.
For example, the level of 2-year yields (gt2 in the �gure) when lagged 15 trading days has
a correlation of -97.4% with the corona series. All other lags of gt2 have a correlation that
is smaller, in absolute terms. The two series are plotted in the upper left-hand corner of
Figure 4. The corona series is shown as a dashed line, and the 15-day lag of gt2, shown
as the solid line, is scaled to have the same range as corona and then multiplied by the
sign of its correlation to corona. 2-year Treasury yields begin falling 15-days prior to the
sharp upturn in COVID-19 case counts and the scaled, lagged gt2 series closely tracks the
daily incidence of COVID-19 case counts.
The �gure shows the markets and text series sorted from highest correlation in the
upper-left corner to lowest correlation in the bottom-right corner. Of the top seven most
correlated series, �ve are the markets series, and the remaining two are s_central bank and
s_credit1, the topical sentiment for the central bank and credit1 topics respectively. The
average optimal lag for the markets series is 15.4 trading days. Thus �nancial markets
appear to have anticipated the increased incidence of COVID-19 case counts by roughly
three calendar weeks. Of the remaining text series there are several whose optimal lags
86C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
0
50
100
gt2lag=15 corr=-97.4
0
50
100
sp500lag=14 corr=-95.3
0
50
100
hylag=11 corr=-95.2
0
50
100
s_central banklag=17 corr=-94.4
0
50
100
vixlag=17 corr=94.3
0
50
100
s_credit1lag=5 corr=-94.0
0
50
100
gt10lag=20 corr=-93.7
0
50
100
countlag=15 corr=91.8
0
50
100
s_creditlag=1 corr=-90.9
0
50
100
s_corp & govt USlag=16 corr=-87.8
0
50
100
s_sportslag=21 corr=-84.2
0
50
100
s_currencylag=2 corr=80.6
0
50
100
s_corp futurelag=14 corr=77.0
0
50
100
s_marketslag=4 corr=76.3
0
50
100
s_corp actuallag=16 corr=-74.8
0
50
100
sent_sdlag=21 corr=64.9
0
50
100
sentlag=20 corr=-47.8
0
50
100
s_europelag=0 corr=43.6
0
50
100
s_healthlag=1 corr=28.4
0
50
100
s_oil & commlag=0 corr=13.6
Scaled news series Corona counts
COVID-19 case counts and market and news variables at optimal lagsDate range of analysis: 2020-01-17 to 2020-04-30
Fig. 4. This �gure shows correlations between daily COVID-19 case counts and themarkets and news series at optimal lags. For each candidate series, a lag is chosen from 0to 21 trading days to maximize the correlation between daily COVID-19 case counts andthe lagged market or news series. The series are plotted in order of highest (upper left)to lowest (lower right) absolute correlations. The optimal correlation and lag are shownfor each series. The lagged market or text series is scaled to have the same range as thecase count series corona and then multiplied by -1 (+1) if the correlation between it andcorona is negative (positive). The scaled markets or text series are shown as solid, bluelines in the �gure. The corona series is shown in each subplot as the dashed, orange line.The y-axis corresponds to units of the corona series, in the thousands.
87C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
have a relatively high correlation with corona, though generally the lags are shorter than
15.4 days. Some news series, for example s_oil & comm and, surprisingly, s_health, have
very low correlations with future disease incidence. Interestingly, aggregate sentiment
Sentt (sent in Figure 4) has a relatively low correlation with future case counts, suggesting
that the decomposition of sentiment into topical categories conveys important information.
The general takeaways of this analysis are: (1) markets are highly anticipatory of
future COVID-19 incidence, and (2) some of the news series are as well, though less so.
Given that markets and news anticipate future coronavirus outcomes, it is natural to
examine how the two interact. I now turn to this analysis.
4 Contemporaneous relationships
I assume the contemporaneous relationship between markets and news is captured by
ht+1 = c+ b>wt+1 + et+1, (3)
where ht+1 is the asset return from day t to t + 1, c is a constant, wt+1 is a k × 1 vector
derived from news �ow and other information available to the econometrician, b is a k×1
constant vector, and et+1 consists of unobservable information. This contemporaneous
markets-news speci�cation has been used in the literature, notably in Tetlock et al. (2008)
and in Glasserman, Li, and Mamaysky (2020) for returns and in Boudoukh et al. (2018)
for squared returns. Nevertheless a theoretical justi�cation for (3) has not yet been
established, to the best of my knowledge. In the appendix, I show that such a justi�cation
obtains from the Campbell and Shiller (1988) and Campbell (1991) result that asset
returns can be decomposed into changes in investor beliefs about future dividend growth
and future discount rates. As I show, the information vector wt+1 can then be interpreted
as innovations to state variables that are useful for forecasting future cash �ows or discount
rates. In the rest of the paper, I do not formally separate the contribution of wt+1 to one
of these two channels, though that is an interesting question for future work.
4.1 Empirical implementation
I specialize (3) as follows:
ht+1 = c+ b1ht + b2ht−1 + b3Nt+1 + b4Nt+1(V IX10t − V IX
10) + b5V IX
10t + et+1. (4)
88C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
The observable information includes two lagged returns, ht and ht−1, which control for the
dependence of current news �ow on past returns, and for the auto-correlation properties of
the dependent variable. None of the results change materially if ht and ht−1 are excluded.
In addition wt+1 contains Nt+1, which can be one of: aggregate daily sentiment Sentt,
the daily standard deviation of article-level sentiment sent_sd, one of the twelve topical
sentiment series Sentt,k, or the COVID-19 case count series corona, which tests the degree
to which markets react to the most recently announced COVID-19 case counts. I refer
to these �fteen variables as the news series in the remainder of the paper, expanding the
term �news� to refer also to COVID-19 case counts. All standard errors use Newey-West
with three lags.
The theoretical analysis in Glasserman, Mamaysky, and Shen (2020) suggests infor-
mation shocks can push the economy into a high-information production, high-volatility
regime, where asset prices become both depressed and extremely sensitive to changes in
fundamentals. Section 6 discusses this mechanism in more detail. The Nt+1(V IX10t −
V IX10
) term tests if the e�ect of Nt+1 on returns di�ers depending on whether mar-
kets are currently in a high- or low-volatility period. The volatility regime is proxied by
V IX10t , which is the day t level of the rolling 10 trading-day average of the VIX index,
and V IX10is the average of V IX10
t in the time period over which the regression is being
estimated. I use the 10-day average of the VIX to smooth out high-frequency variation;
and I use a one-day lag to avoid endogeneity issues.
For the SP500 and HY indexes, ht+1 equals the day t to day t+1 total return. For the
VIX, ht+1 measures the day-over-day di�erence in the VIX index. At high frequency, this
is very similar to the return of investing in a VIX futures contract. For 2- and 10-year
Treasuries, ht+1 is the day-over-day change in yields. Assuming Treasury duration did not
change materially during the sample period, which is reasonable given the relatively low
starting and ending interest rate levels from January to April of 2020, Treasury returns
are to a �rst-order linear in yield changes. Therefore, for all asset classes, returns either
equal ht+1 or are approximately linear in ht+1; I will refer to ht as returns.
For each of the �ve dependent variable, I therefore run �fteen di�erent version of (4),
one for each of the possible Nt+1's. I don't include multiple Nt+1's in the speci�cation at
once because given the small sample size, coe�cient estimates would become unreliable.
The results of this analysis are described below. I discuss some econometric issues
with the present analysis, as well as robustness tests, in Section A.2 of the appendix.
89C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
4.2 Structural break tests
A concern with running the regression in (4) over the entire January through April sample
is that the markets-news relationship may change over time. Markets in February and
March of 2020 were extraordinarily stressed, and perhaps the markets-news relationship
that prevailed in this period of extreme stress did not last the entire duration of even this
short sample. If this is true, then running any analysis over the full period risks simply
averaging across two very di�erent regimes. To test for this possibility, I check whether
there is evidence of a structural break in any of the 75 markets�news pairs that I analyze
(�ve markets and �fteen news series for each market).
When testing for a structural break at a known break point 0 < t < T , one can
calculate the Chow test statistic given by
φ(t) =(SSR− SSRe − SSRl)/k
(SSRe + SSRl)/(Ne +Nl − 2k), (5)
where e (early) refers to dates prior to or equal to t and l (late) refers to those dates
after t, SSR refers to the sum of squared residuals over the entire sample, SSRe (SSRl)
refers to the sum of squared residuals over the early (late) part of the sample, Ne (Nl)
refers to the number of observations in the early (late) part of the sample, and k refers
to the number of regressors. As the number of observations grows, kφ(t) approaches a χ2
distribution with k degrees of freedom. Andrews (1993, 2003) tabulates the distribution
of the maximal kφ(t) over all possible ts in some interval of the data. Speci�cally, for an
unknown break point π which is drawn from an interval [π0, 1− π0] ⊂ (0, 1) of the data,
Andrews (1993,2003) tabulates the distribution of
supπ∈[π0,1−π0]
kφ(πT ). (6)
under the null hypothesis that there is no structural break. The tabulated distribution
depends on k, which here equals 6, and π0 which here is set to 0.3. Note the choice of time
period over which to estimate V IX10
a�ects only the b3 coe�cient in (4) but leaves the
residuals of the regression unchanged; therefore φ(t) in (5) is una�ected by this choice.
For each of the �ve market variables, Figure 5 shows a histogram of the maximal break
points t = πT that maximize (6) for the �fteen speci�cations that I analyze. In addition,
for every break point, I tabulate the number of market�news pairs that are signi�cant
at least at the 10% level according to the distribution tables in Andrews (2003). The
bottom-right panel of the �gure shows the distribution of the break points for all 75 tests
90C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Break dates using all articles
03-0
403
-09
03-1
9
03-2
30
5
10
hy tests=15 sig breaks=14countsignif
03-0
4
03-1
203
-13
03-2
3
0
5
10
sp500 tests=15 sig breaks=9countsignif
03-0
9
03-1
3
03-1
703
-18
03-2
0
0
2
4
6
gt10 tests=15 sig breaks=15countsignif
03-0
9
03-1
2
0
5
10
gt2 tests=15 sig breaks=15countsignif
03-1
103
-12
03-1
303
-16
03-1
7
0
5
10vix tests=15 sig breaks=15
countsignif
03-0
403
-09
03-1
103
-12
03-1
303
-16
03-1
703
-18
03-1
903
-20
03-2
3
0
10
20
all tests=75 sig breaks=68countsignif
Break point dates =[0.3,0.7]: all articles
Fig. 5. For each market variable, there are 15 regressions which are tested for a break:overall sentiment, the standard deviation of sentiment, the twelve topical sentiment series,and the COVID-19 case counts series. The data start on January 17, 2020. The startsshow number of break points that are signi�cant at the 10% level or better using theAndrews (2003) distribution for the maximal Chow statistic with π0 = 0.3.
91C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
conducted, as well as the number of break points that are signi�cant.
All �ve asset classes exhibit strong evidence of a regime break sometime in March.
For the VIX and the two Treasury series, all 15 news variables tested show evidence of a
signi�cant break. For HY 14 of 15 do, and for the SP500 9 out of 15 tests show evidence
of a signi�cant break. The break dates are concentrated on March 9th, March 16th and
17th, and March 23rd. March 16th and 17th are the Monday and Tuesday that follow
an emergency rate cut announced by the Fed on Sunday, March 15th. And the March
23rd break date follows another emergency Fed meeting on Sunday, March 22nd when the
Fed removed quantitative guidance from its announced emergency programs and simply
vowed to purchase Treasury and mortgage-backed securities �in the amounts needed.� The
markets reacted favorably to the Fed's, and other central banks', announcements which
were perceived to be expansive and bold (see Hartley and Rebucci 2020).
The structural break tests suggest that the markets�news relationships that had pre-
vailed in the early parts of the crisis changed around the middle of March. Going forward,
I use Sunday, March 15th as a cuto� date between the early- and late-parts of the sample
period, and run all of my analyses separately in these two subsamples. Allowing the break
between the early and late parts of the sample to equal, for each market�news pair, the
optimal break from (6) produces even stronger results. For this reason, my choice of a
�xed break point for all markets�news pairs is conservative.
4.3 The early subperiod
Table 3 summarizes the results of estimating (4) in the early (before March 15) part of the
sample. The column groupings correspond to a particular market variable, and the rows
correspond to one of the �fteen di�erent news variables. b3 (column EV) is normalized
to report the e�ect of a one standard deviation change in the news variable in units of
standard deviations of the market variable; b4 (column EV*VIXl1) shows the change in
the normalized b3 for a one unit increase in the lagged V IX10. For example, the value of
-0.701 for the e�ect of aggregate sentiment on the VIX means when aggregate sentiment
increases by one standard deviation, the VIX experiences a contemporaneous decline of
0.701 standard deviations of daily changes. I will discuss the EVl1 column in Section 5.
92C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 3
Summary of the contemporaneous and lead-lag model results for daily market changes or returns. The column groupingscorrespond to di�erent market variables, and the rows correspond to the �fteen di�erent explanatory variables. The �rsttwo entries for every market variable show the b3 (EV, for explanatory variable) and b4 (EV*VIX) coe�cients in (4) that aresigni�cant at the 10% level or better. EV column shows the impact of a one standard deviation change in the explanatoryvariable in units of standard deviation of the market variable. EV*VIXl1 column shows the impact of a unit increase inV IX10 on the value of EV. The last entry for each market variable indicates the the c3 (EVl1) coe�cient from (7) that aresigni�cant at the 10% level or better, where (7) is run using the lagged all-article text measures. The summary statisticsunderneath the table are as follows: Jnt sig (joint signi�cance) is the number of speci�cations where b3 and b4 are bothsigni�cant; Hyper (hypersensitivity) is the number of times that both b3 and b4 are signi�cant and have the same sign; Lead(lead-lag relationship) is the number of times that c3 is signi�cant; Under (underreaction) is the number of times b3 and c3have the same sign; Over (overreaction) is the number of times b3 and c3 have opposite signs; and Over+Hyper (overreactionand hypersensitivity) is the number of time b3 and c3 have opposite signs while b3 and b4 have the same sign.
Summary of analysis for the all-article corpus: early subsample
SP500 VIX HY GT2 GT10EV EV*VIXl1 EVl1 EV EV*VIXl1 EVl1 EV EV*VIXl1 EVl1 EV EV*VIXl1 EVl1 EV EV*VIXl1 EVl1
sent 0.761*** 0.085*** −0.701*** −0.081*** 0.603*** 0.104*** −0.347** 0.616*** 0.093** 0.574*** 0.092* −0.218*sent_sd −0.742** −0.106** 1.137*** 0.857*** 0.117*** −1.098*** −1.100** −0.180*** 1.181*** 0.906*** 1.231***sports 0.940** −0.613* −1.041*** 0.645** 0.035* −1.333*** 1.129** 0.891*central bank 0.615*markets 0.857*** 0.075*** −0.798*** −0.069*** 1.079*** 0.110*** −0.442* 0.426* 0.078*** −0.892*** 0.960*** 0.139*** −1.077**health 1.053*** 0.104** −1.099*** −0.113*** 1.245*** 0.171*** −0.892*** 0.940*europe 0.985*** 0.078* −0.829*** 0.676*** 0.101*** −0.617**oil&comm 0.663*** 0.215*** −0.460*** −0.187*** 1.073*** 0.223*** 0.817*** 0.169*** −0.808*** 1.110*** 0.232*** −0.863***currency 0.771** −0.681** 1.262*** 0.148*** 1.127*** 0.139***credit 0.576* 0.061*corp&govt US 0.338** 0.379* −0.077**corp actual 0.297** −0.403**corp future 0.531** 0.089*** 0.457*** 0.095***credit1corona −0.187** −0.083*** 0.830*** 0.163* 0.079*** −0.782*** −0.222*** −0.076*** 0.209* 0.403** 0.519***
Jnt sig = 32, Hyper = 32, Lead = 24, Under = 1, Over = 16, Over+Hyper = 15
93C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
All b3 coe�cients for sentiment series in speci�cations involving the SP500, HY, and 2-
and 10-year Treasury yields that are signi�cant have positive signs. So negative news as
conveyed by either aggregate sentiment or any of the topical sentiment series is associated
with a contemporaneous decline in the SP500 and HY indexes, and with a contempora-
neous decrease in Treasury yields. For the SP500 and HY, the sign of b3 for sent_sd is
negative suggesting a contemporaneous market drop associated with a higher standard
deviation of article-level sentiment. For VIX, all of these signs are reversed. Bad news,
as proxied by aggregate or topical sentiment, or a high sent_sd, is associated with a con-
temporaneous increase in the VIX. Finally, the SP500 and HY indexes fall on days when
the prior day's COVID-19 case count (the corona series) is high. And the VIX rises on
such days. As mentioned earlier, the prior day's case counts come out after that day's
market close, and thus markets react to these with a one-day lag.
Another striking feature of Table 3 is that in all 32 cases when both b3 and b4 are
signi�cant, they have the same sign. On high volatility days, measured by an elevated
level of the 10-day average VIX on the prior day, the e�ects of all news measures are larger
than they are during normal-volatility days. Interestingly, the b4 coe�cient for the corona
series also has the same sign as b3, suggesting that the SP500, VIX and HY indexes all
react more to the same increase in case counts during high-volatility than low volatility
days. I refer to these phenomena as hypersensitivity to contemporaneous news: A portion
of the high volatility that asset markets experience in high-volatility states is not due to
an increased volatility of news �ow, but results instead from an increased sensitivity of
markets to similarly volatile news. This is the prediction of Glasserman, Mamaysky, and
Shen (2020) for markets in the high-information, high-volatility regime.
Tables 6 � 10 in the appendix show the full regression estimates of the model in (4) for
any speci�cation with a signi�cant b3 or b4 coe�cient. The rows of the table correspond to
model estimates for a particular news variable. Note that the large negative b1 coe�cients
for the SP500 and VIX indexes in Tables 6 and 7 respectively are consistent with the large
negative AR(1) coe�cients for these series documented in Table 1. Furthermore, the R2s
in these tables are very high, but should be interpreted with caution because of the small
sample size and persistent regressors. The break date column in these tables corresponds
to the date t∗ which maximizes (6) and the associated kφ(t∗) value is shown in square
brackets. The stars indicate signi�cance relative to the Andrews (2003) tabulated values
for the maximal Chow statistic.
4.4 The late subperiod
94C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 4
Summary of the contemporaneous and lead-lag model results for daily market changes or returns. The column groupingscorrespond to di�erent market variables, and the rows correspond to the �fteen di�erent explanatory variables. The �rsttwo entries for every market variable show the b3 (EV, for explanatory variable) and b4 (EV*VIX) coe�cients in (4) that aresigni�cant at the 10% level or better. EV column shows the impact of a one standard deviation change in the explanatoryvariable in units of standard deviation of the market variable. EV*VIXl1 column shows the impact of a unit increase inV IX10 on the value of EV. The last entry for each market variable indicates the the c3 (EVl1) coe�cient from (7) that aresigni�cant at the 10% level or better, where (7) is run using the lagged all-article text measures. The summary statisticsunderneath the table are as follows: Jnt sig (joint signi�cance) is the number of speci�cations where b3 and b4 are bothsigni�cant; Hyper (hypersensitivity) is the number of times that both b3 and b4 are signi�cant and have the same sign; Lead(lead-lag relationship) is the number of times that c3 is signi�cant; Under (underreaction) is the number of times b3 and c3have the same sign; Over (overreaction) is the number of times b3 and c3 have opposite signs; and Over+Hyper (overreactionand hypersensitivity) is the number of time b3 and c3 have opposite signs while b3 and b4 have the same sign.
Summary of analysis for the all-article corpus: late subsample
SP500 VIX HY GT2 GT10EV EV*VIXl1 EVl1 EV EV*VIXl1 EVl1 EV EV*VIXl1 EVl1 EV EV*VIXl1 EVl1 EV EV*VIXl1 EVl1
sent 0.750*** 0.044** 0.184** 0.804***sent_sd −0.246* 0.023** 0.200** 0.407** −0.188* 0.037*** 0.246**sports 0.259** −0.387*** 0.037* 0.437***central bank 0.677*** 0.034*** 0.215*** 0.675*** 0.026**markets 0.507*** −0.268* 0.245*** 0.518***health 0.481*** 0.039** 0.447***europe 0.700*** 0.050*** 0.364** −0.369** 0.044* 0.301*** 0.652*** 0.036** 0.482***oil&comm −0.305** −0.217** −0.219**currency −0.304* 0.340***credit −0.274** 0.046*** 0.450** −0.044** 0.043*** −0.182***corp&govt US 0.215*** 0.280** 0.142* −0.339**corp actual 0.204**corp future 0.018* 0.176** 0.024**credit1 −0.275*** 0.450*** −0.039** −0.233*corona 0.348*** −0.400** 0.053** 0.390*** −0.038* 0.215**
Jnt sig = 12, Hyper = 4, Lead = 16, Under = 4, Over = 4, Over+Hyper = 0
95C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 4 summarizes the results of estimating (4) in the late (after March 15) part of
the sample. Tables A1 � A5 in the Internet Appendix show the full regression results for
all explanatory variables where at least one of the b3 and b4 coe�cients is signi�cant. The
patterns in the late subsample are starkly di�erent from those in the early subsample.
Of the 75 (�ve markets times �fteen news variables) combinations, there are now only
12 times when b3 and b4 are jointly signi�cant. In only four of these cases do b3 and b4
have the same sign. Compare these numbers to the 32 jointly signi�cant market-news
pairs, of which all 32 have the same sign, in the early part of the sample. Furthermore,
only two news variable come in signi�cant for either of the Treasury series. When it is
signi�cant, topical sentiment in the late subsample generally has the same sign as in the
early subsample, with positive news associated with contemporaneous increases in SP500
and HY, and a drop in the VIX. The credit and credit1 series are an exception. However,
the sign of the b3 coe�cient in the corona speci�cation is opposite in the late subsample
compared to the early one; in the late subsample higher case counts are associated with
positive SP500 and HY returns, and with drops in the VIX.
Overall, the late subsample looks very di�erent from the early one. Markets are no
longer as responsive to news series, and markets are rarely hypersensitive. The mid-March
structural break in the markets-coronavirus-news relationship is economically important,
and occurs right around the time of major policy actions by the Fed. Though the timing
may be coincidental, I speculate that the Fed may have e�ected the structural break in
the markets-news relationship. This suggests a new role for central banks � they might
be uniquely positioned to in�uence the markets-news interaction, and to nudge markets
out of their hypersensitive phase.
4.5 Variance decomposition
It is hard to verify how much of market volatility is due to hypersensitivity. Instead, I
analyze the following counterfactual. I reestimate (4) but after removing the hypersensi-
tivity term, i.e. Nt+1(V IX10t −V IX
10). I refer to this as the restricted speci�cation. The
di�erence in the adjusted R2 of the baseline speci�cation versus that of the restricted spec-
i�cation captures the contribution of the hypersensitivity term to the explained variance
of SP500, HY, VIX or Treasury returns. In a counterfactual where the hypersensitivity
e�ect is absent, presumably this di�erence in adjusted R2s measures how much lower
return variance would be. Tables 6�10 (for the early subsample) and Internet Appendix
Tables A1�A5 (for the late subsample) show this R2 di�erence in the Chg R2 column.
Figure 6 summarizes the results of this analysis for the early subsample. Each panel
96C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
R2 summaries for contemporaneous regressions: All-article, early subsample
0.0 0.2 0.4sent
sent_sds_sports
s_marketss_health
s_europes_oil & comm
s_currencycorona
sp500 avg(Chg R2)=0.13
partr2diffr2
0.0 0.2 0.4 0.6sent
sent_sds_sports
s_marketss_health
s_europes_oil & comm
s_currencycorona
vix avg(Chg R2)=0.12partr2diffr2
0.0 0.2 0.4 0.6sent
sent_sds_sports
s_marketss_health
s_europes_oil & comm
s_corp & govt UScorona
hy avg(Chg R2)=0.20
partr2diffr2
0.0 0.2 0.4
sents_sports
s_central banks_markets
s_oil & comms_currency
s_credits_corp future
gt2 avg(Chg R2)=0.14partr2diffr2
0.0 0.2 0.4 0.6
sents_sports
s_marketss_health
s_oil & comms_currency
s_corp & govt USs_corp future
gt10 avg(Chg R2)=0.23partr2diffr2
0.0 0.2 0.4corona
s_central banks_corp & govt US
s_corp futures_credit
s_currencys_europes_health
s_marketss_oil & comm
s_sportssent
sent_sdall avg(Chg R2)=0.14
partr2diffr2
R2 and change in R2 due to VIX interaction for early subperiod
Fig. 6. The �gures summarize the R2s of the contemporaneous regressions in (4), usingthe all-article text measures for the early subsample. The panels show R2s for each marketvariable for which at least one of the b3 or br coe�cients in (4) is signi�cant at the 10%level or better. Some market variables may be missing in some speci�cations if there wereno signi�cant b3s or b4s for any explanatory variable. Each panel shows the adjusted R2
(labeled partr2) for a version of (4) that excludes the VIX-explanatory variable interaction,i.e. for which b4 = 0, and the di�erence between this and the adjusted R2 of the full model(labeled di�r2). The panel labeled all reports the average of these two measures acrossall market variables.
97C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
shows the change in adjusted R2s for each of the market-news pairs where either the the
news variable or its V IX10 interaction (the b3 or b4 coe�cients in equation 4) is signi�cant
at the 10% level or better. The blue, striped bars show the adjusted R2 of the restricted
speci�cation, and the orange bars show the change in R2 when moving to the baseline
(unconstrained) model. The bottom-right panel shows the R2 decomposition, averaged
across all markets-news pairs. Each panel is labeled with the average increase in R2s when
adding back in the hypersensitivity variable to the contemporaneous regression.
For example, for SP500 returns and the s_oil & comm explanatory variable, the
restricted R2 is 12%, and this increases by 41.5% when the Nt+1(V IX10t − V IX
10) term
is added back in (Table 6 shows the exact numbers). The counterfactual is that the
variance of daily SP500 returns would be 41.5% lower if the hypersensitivity e�ect was
absent. The average across all signi�cant news variables for the SP500 is 13%, and
the average hypersensitivity term contribution across all signi�cant speci�cations is 14%
(bottom right panel).
Figure 7 shows the results of this analysis for the late subperiod. First, there are no
signi�cant instances of news variables or their V IX10 interactions for 10-year Treasury
yield changes, which is absent from the �gure. Second, for the SP500, HY, VIX, and
2-year Treasury series, the average R2 contribution of the hypersensitivity term is very
small. Across all three markets the average increase is only 3%, versus the 14% in the
early subsample.
A version of this analysis that looks at all �fteen news variables for each market series
reaches the same qualitative conclusions. There is strong evidence that hypersensitivity
contributes to the variance of market returns in the early subsample (average R2 increase
of 9%), and almost no such evidence in the late subsample (average R2 increase of 1%).
Of course, counterfactuals, especially in the absence of a structural model, are highly
speculative. But the di�erence between the R2 contribution of the hypersensitivity term
in the early (14%) and late (3%) subsamples does suggest that some portion of the early
subperiod volatility was due to the presence of the hypersensitivity e�ect.
5 Lead-lag relationships
In addition to analyzing the contemporaneous relationship between news �ow and mar-
kets, it is natural to analyze the lead-lag relationship between the two. Do markets lead
news? Or do news lead markets? To gain insight into these questions, I conduct a series
of Granger causality tests. The general speci�cation of these tests mirrors the contempo-
98C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
R2 summaries for contemporaneous regressions: All-article, late subsample
0.0 0.2 0.4 0.6sent
sent_sds_sports
s_central banks_markets
s_healths_europe
s_credits_corp future
s_credit1corona
sp500 avg(Chg R2)=0.03partr2diffr2
0.0 0.2 0.4sent
sent_sds_sports
s_marketss_health
s_europes_currency
s_credits_corp & govt US
s_credit1corona
vix avg(Chg R2)=0.06partr2diffr2
0.0 0.2 0.4 0.6sent
sent_sds_sports
s_central banks_markets
s_healths_europe
s_currencys_credit
s_corp & govt USs_corp future
s_credit1corona
hy avg(Chg R2)=0.02partr2diffr2
0.00 0.05 0.10 0.15 0.20
s_oil & comm
s_corp & govt US
gt2 avg(Chg R2)=-0.02partr2diffr2
0.0 0.2 0.4 0.6corona
s_central banks_corp & govt US
s_corp futures_credit
s_credit1s_currency
s_europes_health
s_marketss_oil & comm
s_sportssent
sent_sdall avg(Chg R2)=0.03
partr2diffr2
R2 and change in R2 due to VIX interaction for late subperiod
Fig. 7. The �gures summarize the R2s of the contemporaneous regressions in (4), usingthe all-article text measures for the late subsample. The panels show R2s for each marketvariable for which at least one of the b3 or br coe�cients in (4) is signi�cant at the 10%level or better. Some market variables may be missing in some speci�cations if there wereno signi�cant b3s or b4s for any explanatory variable. Each panel shows the adjusted R2
(labeled partr2) for a version of (4) that excludes the VIX-explanatory variable interaction,i.e. for which b4 = 0, and the di�erence between this and the adjusted R2 of the full model(labeled di�r2). The panel labeled all reports the average of these two measures acrossall market variables.
99C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
raneous regression in (4):
ρt+1 = c0 + c1ρt + c2ρt−1 + c3τt + c4τt(V IX10t − V IX
10) + c5V IX
10t + et+1. (7)
Here ρt+1 is the day t+ 1 response variable, and τt is the day t test variable. The two lags
of the response variable control for contemporaneous correlations with the test variable,
and for the possibility that the test variable is itself Granger caused by the lagged response
variable. I say that τ Granger causes ρ if the c3 coe�cient above is signi�cant at the 10%
level or better. Standard errors are calculated using Newey-West with three lags.
5.1 Early subsample
I �rst look at the early subsample (dates prior to March 15). Table 11 in the appendix
shows the results of estimating (7) with daily changes in the �ve market series as the test
variable, and the next day's change in the �fteen news series as the response variables. For
each market variable, the columns of the table show the c3 and c4 coe�cient estimates
from (7) for each future news variable. c3 and c4 are reported in standard deviations
of the response variable per a standard deviation change in the test variable; c4 reports
the change in c3 for a unit change in V IX10. Figure 8 shows the Granger causality
relationships from the table as a graph. Every signi�cant link from ht (time t market
variable) to Nt+1 (time t + 1 news variable or corona value) is shown as an arrow. The
arrow is a blue, solid (red, dashed) line when the c3 coe�cient in (7) is signi�cant and
positive (negative).
For example, there is a blue line from SP500 to sent because the one-day lagged SP500
return positively a�ects the next day's aggregate sentiment sent. On the other hand, there
is a red, dashed line from the lagged VIX change to aggregate sentiment, because a day
t increase in the VIX causes a day t+ 1 decline in aggregate sentiment. In all cases, day
t increases in the SP500 and HY indexes, and in the two Treasury series, Granger cause
positive day t + 1 topical sentiment increases. But day t increases in the VIX forecast
next day decreases in all sentiment series, except for sent_sd; not surprisingly, a high VIX
today forecasts a high sent_sd or standard deviation of article-level sentiment tomorrow.
No other market variable Granger causes sent_sd. Also no market variable Granger causes
the COVID-19 case count series corona at a one-day lag, though as shown in Section 3.1
all market series are highly correlated with case counts two to four weeks in the future.
In the markets to news variables tests, there are 30 instances where c3 and c4 from (7)
are both signi�cant; in none of these cases do they have the same sign. The market e�ect
100C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Granger causality network: Markets to news, early subperiod
d_sp500
d_vix
d_hy
d_gt2
d_gt10
sent
sent_sd
s_sports
s_central bank
s_markets
s_health
s_europe
s_oil & comm
s_currency
s_credit
s_corp & govt US
s_corp actual
s_corp future
s_credit1
corona
Granger causality network 2020-01-17 to 2020-03-15Num joint significant = 30 Num same sign = 0
Fig. 8. Granger causality tests, using Newey-West standard errors with three lags. Thed_[market] variables refer to daily returns or changes in the particular market series. Alink is shown in the �gure if the lagged test variable is signi�cant at the 10% or betterlevel in equation (7). A blue, solid (red, dashed) line indicates the coe�cient c3 from thetest to the response variable in (7) is positive (negative). The top of the �gure shows thenumber of links in the graph that are associated with a signi�cant c4 coe�cients from (7),labeled Num joint signi�cant. Also shown are the number of times when both c3 when c4are signi�cant and have the same sign, labeled Num same sign.
101C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Granger causality network: News to markets, early subperiod
sent
sent_sd
s_sports
s_central bank
s_markets
s_health
s_europe
s_oil & comm
s_currency
s_credit
s_corp & govt US
s_corp actual
s_corp future
s_credit1
corona
d_sp500
d_vix
d_hy
d_gt2
d_gt10
Granger causality network 2020-01-17 to 2020-03-15Num joint significant = 18 Num same sign = 18
Fig. 9. Granger causality tests, using Newey-West standard errors with three lags. Thed_[market] variables refer to daily returns or changes in the particular market series. Alink is shown in the �gure if the lagged test variable is signi�cant at the 10% or betterlevel in equation (7). A blue, solid (red, dashed) line indicates the coe�cient c3 from thetest to the response variable in (7) is positive (negative). The top of the �gure shows thenumber of links in the graph that are associated with a signi�cant c4 coe�cients from (7),labeled Num joint signi�cant. Also shown are the number of times when both c3 when c4are signi�cant and have the same sign, labeled Num same sign.
102C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
on future news variables is therefore mitigated when the V IX10 is high. This suggests an
information loss in volatile markets � when markets are extremely volatile, the same unit
of market price move becomes less important for future news �ow.
Table 12 in the appendix shows Granger causality tests in the other direction for the
early subsample. Each row of the table corresponds to one of the �fteen time t news
variables, and each column corresponds to one of the �ve market returns at time t + 1.
Figure 9 shows the graph representation of this table. The arrows in the graph have
the same interpretation as before, with blue solid arrows representing signi�cant positive
links, and red dashed arrows representing signi�cant negative links. Granger causality
from the time t news to time t + 1 markets is a less dense graph than the markets to
news one. Generally the signs of these relationships are opposite to what they were in
the contemporaneous regressions. There are many red dashed arrows suggesting that low
time t topical sentiment Granger causes positive time t+ 1 market price action. I return
to this observation momentarily.
Finally, note that there are 18 cases of jointly signi�cant c3 and c4 coe�cients in the
time t news series to time t+ 1 markets Granger causality tests. In all 18 case, c3 and c4
have the same sign. This suggests that hypersensitivity is not only a contemporaneous
phenomenon. When time t V IX10 is high, the e�ect of time t news on time t+ 1 markets
becomes more pronounced. This observation, and the prior one about the di�erence in
signs between the lagged and contemporaneous news-variable-to-markets relationships,
suggest that the lead-lag phenomena of equation (7) are closely related to the contempo-
raneous relationships of equation (4). Again, I return to this momentarily.
5.2 Late subsample
Table A6 in the Internet Appendix shows the results of estimating (7) for the markets
to news Granger causality tests in the late subsample (dates after March 15). Figure 10
shows these results in graph form. The Granger causality network from market prices to
news is very sparse in the late part of the sample. Along with the other changes that
occur post the mid-March structural break, there is relatively little evidence that markets
Granger cause news �ow. There are now 6 cases of jointly signi�cant c3 and c4 coe�cients
in (7), down from 30 in the early subsample, and again none of them have the same sign.
Table A7 in the Internet Appendix shows the results of estimating (7) for the news
to markets Granger causality tests in the late subsample. Figure 11 shows these results
in graph form. The Granger causality network from time t news variables to time t + 1
market responses is again much more sparse than in the early part of the sample. The
103C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Granger causality network: Markets to news, late subperiod
d_sp500
d_vix
d_hy
d_gt2
d_gt10
sent
sent_sd
s_sports
s_central bank
s_markets
s_health
s_europe
s_oil & comm
s_currency
s_credit
s_corp & govt US
s_corp actual
s_corp future
s_credit1
corona
Granger causality network 2020-03-15 to 2020-04-30Num joint significant = 6 Num same sign = 0
Fig. 10. Granger causality tests, using Newey-West standard errors with three lags. Thed_[market] variables refer to daily returns or changes in the particular market series. Alink is shown in the �gure if the lagged test variable is signi�cant at the 10% or betterlevel in equation (7). A blue, solid (red, dashed) line indicates the coe�cient c3 from thetest to the response variable in (7) is positive (negative). The top of the �gure shows thenumber of links in the graph that are associated with a signi�cant c4 coe�cients from (7),labeled Num joint signi�cant. Also shown are the number of times when both c3 when c4are signi�cant and have the same sign, labeled Num same sign.
104C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Granger causality network: News to markets, late subperiod
sent
sent_sd
s_sports
s_central bank
s_markets
s_health
s_europe
s_oil & comm
s_currency
s_credit
s_corp & govt US
s_corp actual
s_corp future
s_credit1
corona
d_sp500
d_vix
d_hy
d_gt2
d_gt10
Granger causality network 2020-03-15 to 2020-04-30Num joint significant = 7 Num same sign = 5
Fig. 11. Granger causality tests, using Newey-West standard errors with three lags. Thed_[market] variables refer to daily returns or changes in the particular market series. Alink is shown in the �gure if the lagged test variable is signi�cant at the 10% or betterlevel in equation (7). A blue, solid (red, dashed) line indicates the coe�cient c3 from thetest to the response variable in (7) is positive (negative). The top of the �gure shows thenumber of links in the graph that are associated with a signi�cant c4 coe�cients from (7),labeled Num joint signi�cant. Also shown are the number of times when both c3 when c4are signi�cant and have the same sign, labeled Num same sign.
105C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
signs of the relationships are mixed, with some instances of positive time t sentiment
Granger causing positive time t+ 1 market reactions, as well as some instances where the
sign changes. There are now seven jointly signi�cant c3 and c4 coe�cients, of which �ve
have the same sign, and this compares to the 18 jointly signi�cant relationships in the
early part of the sample, of which all 18 had the same sign.
5.3 Feedback loops
Figure 8 makes clear that day t market action has a profound e�ect on day t + 1 news
coverage of the impact of coronavirus across a broad range of topical categories. These
results are reminiscent of Garcia (2018), who shows that future New York Times and Wall
Street Journal news are especially responsive to lagged negative stock returns. Given
that the price action during the early subsample is very bearish, the early part of the
coronavirus crisis is an example of a regime when news articles are extraordinarily sensitive
to lagged market performance. The evidence in the late subsample suggests that lead-lag
relationships between markets and news are then much less pronounced than in the early
subsample; in particular, in Figure 10 news �ow becomes much less sensitive to lagged
price action, as Garcia's (2018) results suggest.
A consequence of the dense, early subsample networks of Figures 8 and 9 is the pos-
sibility of feedback loops, from markets to news back to markets, that may exacerbate
volatility. One form of feedback is time t− 1 market action a�ecting time t news, which
in turn a�ects time t markets via (4). A second form is from time t− 1 markets to time t
news to time t + 1 markets. The speci�cation in (7) controls for the possibility that, for
example, time t−1 SP500 returns Granger cause time t markets sentiment, which in turn
Granger causes time t+1 SP500 returns; such direct loops are precluded by including two
lags of the response variable in (7). It is possible, however, that time t markets sentiment,
having been Granger caused by times t− 1 SP500 returns, goes on to Granger cause time
t + 1 HY returns. In this way, tight lead-lag networks might create excessive volatility
unrelated to COVID-19 case incidence. Recall that case incidence, corona, is not Granger
caused by any market series in the early subsample, as seen in Figure 8.
The late subsample networks of Figures 10 and 11 are much less dense, and thus less
susceptible to such feedback loops.
106C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
6 Interpretation
Thus far, I have presented evidence that markets react more strongly to news and COVID-
19 case counts in high volatility periods. Removing this e�ect from contemporaneous
regressions of market returns on news and case counts meaningfully lowers the R2, sug-
gesting a counterfactual where the hypersensitivity e�ect is missing and return variances
are lower. This hypersensitivity e�ect is present in the early subsample, but is largely
missing from the late subsample. Furthermore, in tests of lead-lag relationships between
market and news variables, I �nd the Granger networks to be much more dense in the
early than the late subsample, and speculate that this is more likely to lead to feedback
loops in the early part of the sample.
I now check whether the news series that are associated with contemporaneous market
responses are also the news series that are informative, in the sense that they forecast
future incidence of COVID-19 cases. I then check whether hypersensitivity and the lead-
lag relationships of Section 5 are related. Finally, I propose a theoretical mechanism that
can explain part of my �ndings.
6.1 Hypersensitivity and the information content of news
As I showed in Section 3.1, the market price series are very informative about two to four
week ahead incidence of COVID-19 cases, as are several news series. During the early
phase of the coronavirus crisis, markets were keenly concerned with understanding the
repercussions of COVID-19 incidence for future economic and societal outcomes. There-
fore, it would seem natural that markets should have been particularly responsive to news
series that would prove to be informative about future incidence of COVID-19. Figure 4
shows that central bank, credit1, credit, corp & govt US and sports topical sentiment were
particularly informative for future case counts, with average lead times of twelve trading
days. Were these, therefore, the news topics that were most strongly contemporaneously
related to market price moves in equation (4)?
The early subperiod summary results in Table 3 show credit1 was not signi�cant in
any of the �ve market speci�cations. The central bank and credit series were each sig-
ni�cant once, for 2-year Treasuries. corp & govt US was signi�cant only for high-yield
and the sports series was signi�cant for all �ve markets, and interestingly associated with
hypersensitivity only once. The other news series that �gure prominently in the contempo-
raneous early subsample regressions are: aggregate sentiment sent, article-level sentiment
standard deviation sent_sd, markets, health, europe, oil & comm, and currency. However,
107C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
again looking at Figure 4, with the possible exception of currency, all of these news series
are particularly poor forecasters of future case outcomes. Furthermore, the vast majority
of these news series are associated with hypersensitivity in the early subsample.
The late subsample summary results for equation (4) in Table 4 show that the two
credit topics now enter with signi�cant b3 coe�cients in �ve markets-news pairs, that
sports enters as signi�cant for three market variables, and corp & govt US and central
bank together have signi�cant b3 coe�cients �ve times. Many of the signi�cant news
series from the early sample remain signi�cant here, though without any clear patterns.
Series that are more informative about future COVID-19 incidence play a larger role in
the late subsample, where there is little evidence of hypersensitivity, than in the early one.
Among the informative news series, sports stands out as the one most contemporaneously
related to market action in both subsamples.
In the early subsample, hypersensitivity occurred with regard to news series that were
not informative about future COVID-19 incidence, though admittedly this would not
have been known to market participants at the time. News series that did forecast future
outcomes were less important for contemporaneous market movements during the early
phase of the crisis. Furthermore the more informative (about future cases) news series
in the early subsample tended not to be hypersensitive. It is highly unlikely, therefore,
that hypersensitivity e�ects were due to the high information content of news series in
high volatility periods. In fact, the news series that were associated with hypersensitivity
were not generally informative about future COVID-19 outcomes. To the extent that such
price reaction would prove to be unwarranted, I now check to see whether hypersensitivity
is associated with overreaction.
6.2 Hypersensitivity and overreaction
For each market-news pair, the EV l1 column of Tables 3 and 4 reports all signi�cant (at
the 10% level or better) c3 coe�cients from the Granger causality regressions of future
markets on current news in equation (7), along with the signi�cant b3 or b4 coe�cients
from the contemporaneous regression in (4). There are 24 (16) cases of news variables
Granger causing next-day market variables, as indicated by a signi�cant entry in the
EV l1 column, in the early (late) subsample. These numbers are shown in the captions
underneath the two tables.
I de�ne underreaction as a day t+1 market move in response to day t news that goes in
the same direction as the day t market response, the implication being that the market did
108C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
not react �enough� in day t to day t news.12 In the present context, underreaction would
occur when b3 (EV column) and c3 (EV l1 column) have the same sign. This happens once
in the early subsample, and four times in the late subsample. Three of these instances
involve the HY index, and one each involves the SP500 and 2-year Treasuries. There is
relatively little evidence of underreaction.
Overreaction occurs when the contemporaneous news coe�cient, b3 from (4), and the
lagged news coe�cient, c3 from (7), have opposite signs. This indicates that the time t
market e�ect of a news variable is partially or fully reversed at time t + 1. Note the c3
coe�cient reported in column EV l1 is scaled in the same way as coe�cients b3 and b4 in the
EV and EV ∗V IXl1 columns; it is reported in standard deviations of the market variable
per unit of standard deviation of the news variable. Thus the magnitudes of coe�cients in
the EV (contemporaneous regression) and EV l1 (Granger causality regression) columns
of Tables 3 and 4 are directly comparable.
For example, in the early subsample, sent_sd's e�ect on SP500 contemporaneous re-
turn is a decrease of 0.742 standard deviations of SP500 returns for a one standard de-
viation in increase in sent_sd. The coe�cient of lagged sent_sd for future SP500 returns
suggests a one standard deviation increase in sent_sd forecasts a next-day positive SP500
return of 1.137 standard deviations. This completely reverses the contemporaneous SP500
reaction to the standard deviation article-level sentiment.
Counting instances of b3 and c3 coe�cients with opposite signs indicates that this oc-
curs 16 times in the early subsample, and four times in the late subsample (counts shown
in captions underneath the tables). So while there is little evidence of underreaction of
markets to news, especially in the early subsample, there is strong evidence of overreaction
of markets to news in the early subsample, with somewhat weaker evidence of the same
e�ect in the late subsample. Interestingly, of the 16 instances of overreaction in the early
subsample, 15 occur in the presence of hypersensitivity in the contemporaneous markets-
news relationship. Recall that hypersensitivity in Table 3 occurs when the coe�cients
b3 and b4 from (4), shown in columns EV and EV*VIXl1, have the same sign. The fact
that of the 16 instances of early subsample overreaction, 15 of those instances happen
for markets-news pairs that are hypersensitive suggests that hypersensitivity is associated
with overreaction. Importantly, in Table 3 there is evidence of overreaction and hyper-
sensitivity not only for markets-news pairs, but also for the markets-corona combination.
During high-volatility periods, markets experience excessive sensitivity to COVID-19 case
counts, which then gets reversed in the next trading day. In the late subsample, none of
12Glasserman, Li, and Mamaysky (2020) explores stock-level over- and under-reaction to lagged news.
109C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
the four instances of overreaction is associated with hypersensitivity.
There is little evidence of underreaction during the coronavirus crisis. There is, how-
ever, compelling evidence for overreaction. Furthermore, overreaction appears strongly
associated with hypersensitivity.
6.3 Theoretical mechanism
The theoretical mechanism proposed in Glasserman, Mamaysky and Shen (GMS 2020)
yields important insights into price action around the coronavirus crisis. The paper de-
velops a dynamic, asymmetric information model in the spirit of Grossman and Stiglitz
(1980) where the information state is persistent from one time period to the next. An
application of the model to the coronavirus crisis would require that agents can acquire,
at a cost, information about the pandemic that is not widely available. Given that many
professional investors had extensive conversations with medical experts in the early stages
of the crisis, where access to those experts was not widely available, supports this assump-
tion. Though in reality, as in theory, there is spillover from this news acquisition to prices.
Indeed, as I showed in Section 3.1, market prices have been very informative about future
fundamentals, here measured as COVID-19 case counts.
The model yields an economy where markets can be in either a low- or a high-
information regime. Markets endogenously transition from one regime to the other in
response to information shocks, and each regime is highly persistent.13 In the high-
information regime, a news production sector generates a large volume of news because
it anticipates high investor demand for news. In the high-information state, investors
understand that tomorrow's news production is likely to be very high, and therefore that
tomorrow's prices will be very volatile. To protect against this volatility, investors de-
mand a very high risk premium today, and their high risk sensitivity makes today's prices
highly responsive to today's news �ow. Therefore when markets enter the high volatility
regime, stock prices fall (and risk premia rise), and realized price volatility is very high.
In GMS (2020) such high-volatility, low-price states can arise without any change
in fundamentals; in particular, high market price volatility is not associated with high
fundamental volatility. This is because, in the high information state, prices become
hypersensitive, in the sense that the coe�cients in the linear price function become larger
13Of course, the world is far more complex than what the model can capture, and aspects of themodel are unrealistic. One thing the model gets wrong relative to the present crisis is that the high-information state in the base calibration of the model is much longer-lived than the two-months or so ofthe hypersensitive regime in the early subsample.
110C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
in magnitude and identical supply or dividend shocks have a larger price impact. Clearly,
the coronavirus pandemic is characterized by a large change in fundamentals, but the
extreme sensitivity of market prices to news in high-V IX10 periods suggests one dimension
of the crisis is the market being in a high-information state. News production today
begets news production tomorrow. Prices respond to news as I've shown empirically, and
the fear of high news production tomorrow causes prices today to be low and extremely
sensitive to today's news �ow. The model and empirical results suggest the high volatility
experienced by �nancial markets during the coronavirus pandemic is, at least partially,
caused by extreme market responses to news, rather than by extreme news �ow itself.
7 Conclusion
I have documented that a twelve topic model captures the news narrative about the
coronavirus crisis of 2020, and optimally balances the trade-o� between having a higher
number of topics and topic coherence. Using structural break tests, I show that the con-
temporaneous relationship between market returns and news undergoes a regime break
around the middle of March 2020. In the early subsample, I show that the contemporane-
ous e�ect of news on markets is characterized by a hypersensitive state, where the same
quantum of news causes larger market price reactions on high volatility days. Such hyper-
sensitivity is associated with excessive volatility, as well as with market overreaction to
contemporaneous news, which gets systematically reversed the next trading day. Finally,
while there are several news series that are informative about future fundamentals, as
measured by COVID-19 case incidence, these are not the news series that are most asso-
ciated with contemporaneous price moves. Therefore much of market hypersensitivity to
news is largely unwarranted by future fundamentals. However, market prices themselves
are able to e�ectively forecast two to four week ahead COVID-19 case counts.
Overall, the hypersensitive market state is not bene�cial. It makes prices overly
volatile, and is associated with price overreaction to news. Furthermore, the analysis
in Glasserman, Mamaysky, and Shen (2020) suggests that hypersensitive markets are as-
sociated with low prices and high risk premia, which is also not bene�cial to society as a
persistently high cost of capital would inhibit investment. The evidence of a structural
break in mid-March, right around the time of extensive intervention in markets by the
Fed and other central banks, suggests another role central banks can play in the economy.
They can nudge markets out of hypersensitive states into more normal ones.
111C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
References
Altman, E., 2020, �COVID-19 and the credit cycle,� Journal of Credit Risk, forthcoming.
Andrews, D., 1993, �Tests for parameter instability and structural change with unknown
change point,� Econometrica, 61 (4), 821�856.
Andrews, D., 2003, �Tests for parameter instability and structural change with unknown
change point: A corrigendum,� Econometrica, 71 (1), 395�397.
Baker, S., N. Bloom, S. Davis, K. Kost, M. Sammon and T. Viratyosin, 2020, �The
unprecedented stock market reaction to COVID-19,� Covid Economics, 1, 33�42.
Barclay, M. and T. Hendershott, 2004, �Liquidity externalities and adverse selection:
Evidence from trading after hours,� Journal of Finance, 59 (2), 681�710.
Boudoukh, J., R. Feldman, S. Kogan and M. Richardson, 2018, �Information, trading,
and volatility: Evidence from �rm-speci�c news,� Review of Financial Studies, 32 (3),
992-1033.
Calomiris, C. and H. Mamaysky, �How news and its context drive risk and returns around
the world,� 2019, Journal of Financial Economics, 133 (2), 299�336.
Campbell, J., 1991, �A variance decomposition for stock returns,� The Economic Journal,
101 (405), 157�179.
Campbell, J. and R. Shiller, 1988, �The dividend-price ratio and expectations of future
dividends and discount factors,� Review of Financial Studies, 1 (3), 195�228.
Das, S. and M, Chen, 2007, �Yahoo! for Amazon: Sentiment extraction from small talk
on the web,� Management Science, 53 (9), 1375�1388.
Fair, R., 2002, �Events that shook the market,� Journal of Business, 75 (4), 713�731.
Garcia, D., 2013, �Sentiment during recessions,� Journal of Finance, 68 (3), 1267�1300.
Garcia, D., 2018, �The kinks of �nancial journalism,� working paper.
Glasserman, P., H. Mamaysky, and Y. Shen, 2020, �Dynamic information regimes in
�nancial markets,� working paper.
Glasserman, P., F. Li, and H. Mamaysky, 2020, �Time variation in the news-returns
relationship,� working paper.
Gormsen, N. and R. Koijen, 2020, �Coronavirus: Impact on stock prices and growth
expectations,� working paper.
112C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Grossman, S. and J. Stiglitz, 1980, �On the impossibility of informationally e�cient mar-
kets,� The American Economic Review, 70 (3), 393�408.
Hartley, J. and A. Rebucci, 2020, �An event study of COVID-19 central bank quantitative
easing in advanced and emerging economies,� NBER working paper 27339.
Ke, S., J.L. Montiel Olea, and J. Nesbit, 2020, �A robust machine learning algorithm for
text analysis,� working paper.
Loughran, T. and B. McDonald, 2011, �When is a liability not a liability? Textual analysis,
dictionaries, and 10-Ks,� Journal of Finance, 66, 35�65.
Loughran, T. and B. McDonald, 2020, �Management disclosure of risk factors and COVID-
19,� working paper.
Michal, R.-Z., T. Gri�ths, M. Steyvers, and P. Smyth, 2004, �The author-topic model for
authors and documents,� Proceedings of the 20th conference on Uncertainty in arti�cial
intelligence, 487�494.
Newman, D., J. Han Lau, K. Grieser, and T. Baldwin, 2010, �Automatic evaluation of
topic coherence,� Human Language Technologies: The 2010 Annual Conference of the
North American Chapter of the ACL, 100�108.
Roll, R., 1988, �R2�, Journal of Finance, 43 (3), 541�566.
Steyvers, M. and T. Gri�ths, 2007, �Probabilistic topic models,� in T. K. Landauer, D. S.
McNamara, S. Dennis, and W. Kintsch (Eds.), Handbook of Latent Semantic Analysis,
pp. 427-448, Lawrence Erlbaum Associates Publishers.
Tetlock, P., 2007, �Giving content to investor sentiment: The role of media in the stock
market,� Journal of Finance, 62, 1139�1168.
Tetlock, P., M. Saar-Tsechansky, and S. Macskassy, 2008, �More than words: Quantifying
language to measure �rms' fundamentals,� Journal of Finance, 63 (3), 1437�1467.
A Appendix
A.1 Derivation of equation (3)
Campbell (1991) shows that ht+1 can be approximated as
ht+1 ≈ c+ (Et+1 − Et)∞∑j=0
ρj∆dt+1+j − (Et+1 − Et)∞∑j=1
ρjht+1+j, (8)
113C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
where ρ < 1 is a constant discounting factor, c equals Etht+1 which is assumed to beapproximately constant, dt is the time t log dividend, Et is an expectation taken over theinvestor's information set, and the change-in-beliefs operator (Et+1 − Et)X is shorthandfor Et+1X − EtX.
As in Campbell (1991), assume that an M -dimensional state vector zt follows a VARprocess
zt+1 = Azt + εt+1.
The one-lag speci�cation is without loss of generality since zt can be augmented withlagged state variables. Let the �rst two elements of zt be the dividend growth ∆dt andthe return ht, with the remaining elements representing other information useful for fore-casting future returns and dividends. The change in beliefs from t to t+ 1 about zt+T isgiven by
(Et+1 − Et)zt+T = (AT−1zt+1 − AT zt) = AT−1εt+1.
Therefore ht+1 from (8) can be written as
ht+1 ≈ c+ e1>εt+1 + e2>(ρA+ ρ2A2 + · · · )εt+1 = c+(e1> + e2>ρA(I − ρA)−1
)εt+1 (9)
where I is the M ×M identity matrix, e1 is an M ×1 vector with a 1 in the �rst element,and zeros in all the others, and e2 is an M × 1 vector with a 1 in the �rst element, -1in the second, and zeros everywhere else. The linear combination of εt+1 in (9) can bedecomposed into a part that loads on wt+1, the information set of the econometrician,and an orthogonal part et+1. With this, equation (3) follows.
A.2 Robustness
Sample size and persistence
The present analysis, unavoidably, involves a small number of daily observations. Themain research question is how markets and news interact during the coronavirus crisis, andthe crisis is (hopefully) a short-lived episode. The small sample size is thus a fundamentalconstraint of the setting. An issue is that some of the regressors, for example corona, arevery persistent. All of the paper's regressions use daily data, and none of the observationsare overlapping. The persistence therefore is not mechanical, but rather a time-seriesproperty of the underlying data. To control for this, I use two lags of the dependentvariables in the contemporaneous speci�cation in (4), and also use the same two lagsin the lead-lag speci�cation in (7). I also tried running the regression in (4) in �rstdi�erences, thus reducing persistence in all variables, and the results were not materiallydi�erent. Of course, all standard errors also use the Newey-West HAC estimator.
A mitigating consideration is that I run all the analysis separately for the early- andlate-parts of the sample. The small-sample and persistence concerns are present in bothparts of the sample, and yet the early- and late-subsample results are fundamentallydi�erent. This di�erence supports the assertion that neither set of results obtains fromsmall-sample biases.
Furthermore, I run all the paper's results at the level of 75 markets-news pairs. There
114C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
are important di�erences between results for di�erent market-news pairs, and yet all spec-i�cations are subject to the same econometric issues. The di�erences in markets-newsrelationships across the 75 such combinations (5 market variables and 15 news variables)in the contemporaneous and lead-lag regressions suggests that results are capturing fun-damental aspects of the data, and are not in�uenced by a speci�c bias. Finally, the factthat all 32 jointly signi�cant b3 and b4 coe�cients in the early subsample (Table 3) havethe same sign, and that only four of the 12 jointly signi�cant b3 and b4 coe�cients in thelate subsample do (Table 4), is highly unlikely to stem from small sample issues.
Endogeneity
The contemporaneous relationship in (4) is not necessarily indicative of a causal relation-ship from news to market prices, though this is the likeliest explanation. It is possiblethat news articles are written in response to market prices, in which case news �ow clearlydoes not cause price changes. This concern is most pronounced for topics explicitly dis-cussing market prices, such as the markets topic. However, there are several mitigatingconsiderations. First, other hypersensitive markets-news pairs are explicitly not about�nancial markets. For example, the sports and health topics are concerned with issuesother than market price action, as evidenced by the headlines in Table 5. Second, sent_sdis not in any obvious way related to contemporaneous market price action, and indeedre�ects the dispersion of sentiment across hundreds or thousands of daily articles, most ofwhich are not about market activity. And sent_sd comes in as an important market-newspair for SP500, HY, and VIX. Third, the corona series, which is obviously fully exogenousfrom contemporaneous market prices, also comes in as a signi�cant � and hypersensitive� markets-news pair for the SP500, HY, and VIX indexes.
Furthermore my analysis leans heavily on the interaction of the b3 and b4 coe�cientsfrom (4) and there is no reason why reverse causality would induce the two coe�cientsto always have the same sign in the early subsample, and to never do so in the latesubsample. And the two subsamples are subject to the same endogeneity considerations.
Furthermore, none of the analysis involving the lead-lag speci�cation in (7) is subjectto this concern, because here the explanatory variables are lagged by one-day relative tothe dependent variables. In the news to markets versions of these regressions, in the earlysubsample, out of the 18 cases of signi�cant c3 and c4 coe�cients, the two always havethe same sign (see Figure 9 and Table 12). But in the late subsample, there are only5 news-markets pairs that evidence hypersensitivity (see Figure 11 and Table A7 in theInternet Appendix.) Therefore, hypersensitivity is present in the lead-lag speci�cation in(7) which is free of any endogeneity concerns.
Also the �nding in Section 6.2 that hypersensitivity is closely associated with overre-action does not follow in any mechanical way from potential endogeneity in (4).
Dropping intraday articles
To further address endogeneity concerns with regard to the contemporaneous speci�cationin (4), I rerun all the analysis in the paper but exclude any non-weekend intraday news
115C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
articles. These articles are de�ned as Monday through Friday news that come out between9:30am and 4pm NY time; I refer to the complement of this set as the overnight articles.I recalculate Sentt, sent_sd, and each of the twelve topical sentiment measures using onlythe overnight articles. The topical sentiment series use the same topic model described inSection 2, and which is estimated using all articles in the corpus. Of course, the coronaseries is una�ected by this.
Barclay and Hendershott (2004) show that after hours trading (from 4-6:30pm andthen from 8-9:30am) contains �less than 1/20 as many trades per unit time� as does tradingduring regular market hours. This vastly diminishes the probability that overnight newsstories are written in response to contemporaneous market action. Of course, news storiesthat re�ect past market action are not problematic. Figure 1 shows each of the overnightseries as the orange, dashed line. As can be seen, the overnight version of the series arevery close to the all-article versions. The number of articles, the count series from the�gure, is of course lower for the overnight articles.
The Internet Appendix shows the results of this analysis. Figure A1 shows the struc-tural break tests for the overnight news series. Figures A2 and A3 show the R2 modelsummaries discussed in Section 4.5. Tables A8 and A9 summarize the results of (4) for themarkets-overnight news pairs. Tables A10 � A14 show the results for the early subsampleusing the overnight news measures. Tables A15 � A19 show the results from the sameanalysis in the late subsample. In all cases, these results are qualitatively very similar tothose that use the all-article version of the news series. The small di�erences that ariselikely stem from excluding intraday articles that are not endogenous, but informative. Iconclude endogeneity is unlikely to have an e�ect on the main takeaways of the paper.
A.3 Additional results
116C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 5
Headlines of extreme sentiment (top two most negative and positive) news stories whosedocument-topic weight for a given topic is above 0.70.
Headlines for representative articles by topic
label headline sent Date
sports Rugby-Champions Cup, Challenge Cup quarter-�nals postponed due to coro-navirus
-0.119048 2020-03-16
sports Olympics-Judo quali�ers scrapped through end-April over coronavirus fears -0.108108 2020-03-09sports Rugby-Ireland's Healy to miss rest of Six Nations with hip injury -0.099237 2020-02-26sports Soccer-Barcelona directors quit, throwing club into crisis -0.096774 2020-04-10sports Tom Hanks' son con�dent dad will make full recovery 0.047297 2020-03-12sports UPDATE 1-Soccer-Juventus's Dybala tests positive for coronavirus 0.051095 2020-03-21sports Soccer-Juve's France mid�elder Matuidi tests positive for coronavirus - club 0.062500 2020-03-17sports Soccer-Premier League convenes emergency club meeting after Arteta positive
test0.090909 2020-03-12
central bank Insurers warns on forced payouts for uncovered coronavirus losses -0.169014 2020-04-06central bank Bank of England cancels annual stress test of banks -0.112245 2020-03-20central bank Much more serious than the 2008 crisis, former BoE Governor King says -0.104895 2020-03-23central bank Hungary govt expands tax relief measures, suspends evictions - PM Orban -0.103448 2020-03-23central bank Kuwait central bank ready to take all necessary measures to ensure stability 0.041667 2020-03-30central bank BRIEF-Gulf Central Banks Governors Say Regional Financial Sector Strong
Enough to Face Coronavirus - Wam0.042254 2020-03-16
central bank Swedish central banker tests positive for new coronavirus 0.050000 2020-03-09central bank BRIEF-Canada's move to reduce drop in revenue requirement for businesses
to claim emergency wage subsidy will only be valid in March0.058824 2020-04-08
markets U.S. stock futures tumble at open on coronavirus contagion fears -0.089552 2020-03-08markets BUZZ-Australian tech stocks drop most in over 4 years on virus fears -0.082569 2020-03-08markets Australia shares likely to dip at open, NZ up -0.080000 2020-04-07markets China shares swoon, Hong Kong plummets amid global virus panic -0.078431 2020-03-12markets BUZZ-Australian tech stocks hit record high as U.S. peers surge 0.041237 2020-02-04markets CANADA STOCKS-TSX opens higher on reports of e�ective coronavirus
drugs0.045455 2020-02-05
markets BUZZ-Aussie energy index sees best day in over a month as oil prices jump 0.061224 2020-02-05markets BUZZ-Australia's gold stocks hit near 2-wk peak as virus fears boost safe-
haven demand0.062500 2020-02-10
health Vietnam arrests seven over procurement 'wrongdoings' during coronavirus cri-sis
-0.171429 2020-04-22
health Coronavirus crisis delays opening of Netanyahu trial -0.140625 2020-03-15health Kremlin denies EU allegations Russia is feeding coronavirus disinformation
campaign-0.134615 2020-03-18
health Brazil court OKs investigating allegations Bolsonaro tried to interfere withpolice
-0.134259 2020-04-27
health Britain making good progress with antibody tests - junior minister 0.068966 2020-04-27health PM Johnson fully able to run coronavirus response despite positive test -
spokesman0.069767 2020-03-30
health BRIEF-Positive Coronavirus Case Detected In India's Capital New Delhi-Health Ministry Statement
0.071429 2020-03-02
health UK PM making very good progress in COVID-19 recovery, o�ce says 0.076923 2020-04-11europe BRIEF-Coronavirus Is Not A Major Concern For British Insurer Rsa CEO -0.094340 2020-02-27europe Reuters Insider - Global auto industry threatened by virus crisis -0.092105 2020-02-07europe Reuters Insider - Medic unions say Spain ignored doctors' warnings -0.090909 2020-04-02europe Reuters Insider - Lebanese protesters and soldiers clash in Tripoli -0.090909 2020-04-30
117C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 5
Headlines of extreme sentiment (top two most negative and positive) news stories whosedocument-topic weight for a given topic is above 0.70.
label headline sent Date
europe Reuters Insider - Doctor who showed Putin round hospital tests positive 0.035714 2020-03-31europe Reuters Insider - Reuters Today: European shares open lower, Tesco says
coronavirus costs could top ¿925 million0.036145 2020-04-08
europe LIVE MARKETS-Morning call: Positive thinking 0.038835 2020-04-09europe LIVE MARKETS-Morning call: Clinging to the rebound 0.041379 2020-01-29oil & comm France's Total rejects force majeure notice from Chinese LNG buyer -0.081395 2020-02-06oil & comm UPDATE 1-Speculators cut U.S. crude oil net longs-CFTC -0.080292 2020-01-31oil & comm FACTBOX-U.S. ethanol plants slash output as coronavirus hits fuel demand -0.078740 2020-03-26oil & comm Marathon Petroleum to idle 26,000-bpd Gallup re�nery - source -0.070423 2020-04-08oil & comm CNPC to boost Xinjiang oil�eld daily crude oil output to 36,000 T by end-June 0.019608 2020-03-25oil & comm U.S. wheat futures rise with demand for bread strong 0.020513 2020-03-18oil & comm CBOT wheat closes �rm on strong demand 0.028571 2020-03-20oil & comm GRAINS-Soybeans, wheat climb on currency; corn follows crude oil 0.028736 2020-04-09currency BUZZ-US jobless claims back in focus as warning sign, Philly Fed tumbles -0.121212 2020-03-19currency BUZZ-COMMENT-Euro suppressing factors are now weaker -0.111111 2020-03-05currency FOREX-Dollar erases early losses on global economy worries -0.099237 2020-04-15currency BUZZ-COMMENT-AUD/JPY vulnerable as coronavirus fears escalate -0.098266 2020-03-15currency BUZZ-AUD/USD-Moves higher after better than expected Aus GDP 0.041667 2020-03-03currency BUZZ-COMMENT-It's good to think positively, better to play safe 0.042017 2020-04-02currency REFILE-BUZZ-EUR/USD-Bears maintain their grip despite risk-on 0.046512 2020-04-06currency Turkish lira �rms 0.7% to strongest level in a week 0.068493 2020-03-25credit Fitch Downgrades South African Banks to 'BB', Negative Outlook on Coron-
avirus Impact-0.047221 2020-03-31
credit Fitch Takes Actions on Colombian FIs & Related Entities After SovereignDowngrade
-0.046079 2020-04-08
credit Fitch Takes Action on 14 Italian Banking Groups On Coronavirus Disruption -0.042446 2020-03-24credit Fitch Takes Actions on 13 Mexican Banks Due to Operating Environment
Deterioration-0.036833 2020-04-21
credit Fitch Rates Texas Instruments' $750 Million of Five-Year Senior Notes 'A+' 0.010546 2020-03-03credit Fitch Rates KLA Corp.'s Senior Notes O�ering 'BBB+' 0.011581 2020-02-19credit Fitch A�rms Sri Lanka's Dialog Axiata at 'AAA(lka)'; Outlook Stable 0.012177 2020-03-31credit Fitch A�rms L'Oreal at 'F1+'; Withdraws Ratings 0.012441 2020-02-27corp & govt US BRIEF-HHS Launches Covid-19 Uninsured Program Portal -0.113208 2020-04-27corp & govt US U.S. House Speaker Pelosi calls Trump WHO decision senseless, dangerous -0.111111 2020-04-15corp & govt US Pelosi calls on Trump to speed production of critical medical equipment -0.111111 2020-03-19corp & govt US BRIEF-FDA & FTC Warn 7 Companies Selling Fraudulent Products That
Claim To Treat Or Prevent COVID-19-0.106383 2020-03-09
corp & govt US BRIEF-CEPI Collaborates With The Institut Pasteur In A Consortium ToDevelop Covid-19 Vaccine
0.047619 2020-03-19
corp & govt US BUZZ-Borqs set for best day ever after receiving $150 mln line of credit 0.053763 2020-04-13corp & govt US Washington, D.C., has �rst 'presumptive positive' test for coronavirus 0.055556 2020-03-07corp & govt US BRIEF-GSK Announces Collaboration With CEPI To Develop Coronavirus
Vaccine0.061224 2020-02-02
118C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 5
Headlines of extreme sentiment (top two most negative and positive) news stories whosedocument-topic weight for a given topic is above 0.70.
label headline sent Date
corp actual BRIEF-Galaxy Gaming Suspends Billing During Casino Closures -0.151515 2020-03-16corp actual BRIEF-SilkAir Says To Suspend Operations To Hiroshima -0.147059 2020-03-03corp actual BRIEF-Portland General Electric Says Suspending Disconnections, Late Fees
For Customers During Coronavirus Crisis-0.142857 2020-04-06
corp actual Australia Stock Exchange release from FLETBUILD <FBU.AX>: Covid-19update cancels interim dividend suspends buyback
-0.133333 2020-03-24
corp actual BRIEF-Luna Innovations Provides Business Update On Covid-19 0.108108 2020-04-14corp actual BRIEF-Inspiration Health Says 15% Rev Growth To ¿17.8 Mln In FY 0.123967 2020-04-21corp actual Australia Stock Exchange release from SOMNOMED <SOM.AX>: Satisfac-
tory Q3 despite COVID-19 a�ected March0.142857 2020-04-26
corp actual Australia Stock Exchange release from FRONTIER DIGITAL <FDV.AX>:Improved pro�tability and balance sheet despite COVID-19
0.200000 2020-04-29
corp future UPDATE 1-German investor morale worsens on coronavirus fears - ZEW -0.095000 2020-02-18corp future BUZZ-Hershey: Falls on Q1 pro�t miss, massive sales decline in China -0.091954 2020-04-23corp future U.S. weekly jobless claims surge to a record 3.28 mln as coronavirus spurs mass
layo�s-0.087912 2020-03-26
corp future Italy think-tank sees recession in Q1 over coronavirus -0.084507 2020-02-27corp future BUZZ-Nvidia down slightly after upbeat Susquehanna report 0.053191 2020-03-31corp future BUZZ-Analog Devices: Up as brokerages remain positive on future growth 0.055046 2020-02-20corp future BUZZ-Micron lifts as Raymond James upgrades to "strong buy" on demand
growth0.055172 2020-02-06
corp future BUZZ-Brokerage sees Advanced Micro Devices gaining share from Intel, up-grades AMD
0.057143 2020-03-03
credit1 Fitch Downgrades ClearBridge Funds' Notes to 'A' & MRPS to 'BBB'; RatingWatch Negative
-0.047769 2020-03-17
credit1 Fitch Downgrades Kayne Anderson Closed End Funds' Notes to 'A' & MRPSto 'BBB'; Rating Watch Neg.
-0.043635 2020-03-17
credit1 Fitch Ratings De�nes Coronavirus Scenarios for U.S. Toll Roads -0.040597 2020-03-24credit1 Fitch Ratings: USPF Housing De�nes Coronavirus Scenarios for Loan Pro-
gram Models-0.040170 2020-04-30
credit1 Fitch A�rms Jackson County Schneck Memorial Hospital's (IN) Revs at 'AA-'; Outlook Stable
0.011084 2020-03-20
credit1 Fitch A�rms Franciscan Communities, Inc. (IL) at 'BBB-'; Outlook RemainsPositive
0.013284 2020-03-19
credit1 Fitch A�rms FirstHealth of the Carolinas, NC's Bonds at 'AA'; Outlook Sta-ble
0.014033 2020-03-24
credit1 Fitch A�rms Chubb's 'AA' IFS Ratings; Outlook Stable 0.014760 2020-04-03
119C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 6
Contemporaneous regression of daily changes in SP500 index on text variables and corona.Rows correspond to each explanatory variable for which either the b3 or b4 coe�cient in(4) is signi�cant at the 10% level or better. EV column shows the impact of a one standarddeviation change in the explanatory variable in units of standard deviation of the SP500index. EV*VIXl1 column shows the impact of a unit increase in V IX10 on the value ofEV. The V IXl1 column shows the e�ect of a one standard deviation increase in V IX10 inunits of standard deviation of the market variable. Numbers in parentheses are t-statitics,and the numbers in square brackets underneath the adjusted R2s are joint F-tests. Bothsets of tests use Newey-West standard errors with 3 lags. The R2 column shows adjustedR2s, and Chg R2 shows the di�erence in adjusted R2s between the speci�cation in (4)and a version of (4) which sets b4 = 0. The Break Date column shows the structuralbreak date of (4) using the procedure from Section 4.2, and the maximal Chow statistic isshown in square backets, with sigin�cance levels obtained from the Andrews (1993, 2003)distribution indicated via the *s. Signi�cance at the 10%, 5%, or 1% levels is indicatedvia *,**,*** respectively.
Contemp. changes in SP500 index: early subperiod, VIX interact
const ht ht−1 EV EV*VIXl1 VIXl1 R2 Chg R2 Break Date
sent -23.464*** -0.745*** -0.106 0.761*** 0.085*** 5.595*** 0.521 0.176 2020-03-12(-2.71) (-4.56) (-0.64) (5.72) (4.43) (4.16) [0.000] [22.15**]
sent_sd -41.755 -0.743** -0.215 -0.742** -0.106** 8.795** 0.194 0.061 2020-03-23(-1.60) (-2.27) (-0.87) (-2.21) (-2.24) (2.24) [0.032] [23.94***]
s_sports -3.391 -0.733*** -0.327 0.940** 0.023 0.953*** 0.231 0.002 2020-03-23(-1.34) (-3.16) (-1.42) (2.54) (1.17) (2.81) [0.001] [14.93]
s_markets -0.774 -0.698*** -0.239 0.857*** 0.075*** 1.087** 0.324 0.107 2020-03-23(-0.15) (-3.08) (-0.96) (3.72) (4.85) (2.18) [0.000] [20.64**]
s_health -4.113 -0.693*** -0.321 1.053*** 0.104** 1.151** 0.293 0.062 2020-03-04(-1.35) (-2.82) (-1.33) (2.58) (2.41) (2.10) [0.048] [12.79]
s_europe -10.194 -0.498*** -0.330 0.985*** 0.078* 2.817* 0.390 0.066 2020-03-23(-1.03) (-3.57) (-1.61) (5.03) (1.86) (1.78) [0.000] [18.43*]
s_oil & comm -29.775*** -0.321* 0.011 0.663*** 0.215*** 4.728*** 0.536 0.415 2020-03-23(-4.11) (-1.94) (0.06) (4.05) (4.83) (4.53) [0.000] [33.80***]
s_currency 2.197 -0.541*** -0.088 0.771** 0.049 0.499 0.238 0.009 2020-03-23(0.46) (-3.31) (-0.32) (2.48) (1.05) (0.60) [0.001] [13.51]
corona -6.182*** -0.428*** -0.048 -0.187** -0.083*** 0.883*** 0.419 0.314 2020-03-13(-2.98) (-2.92) (-0.39) (-2.03) (-7.09) (2.98) [0.000] [21.50**]
120C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 7
Contemporaneous regression of daily changes in VIX index on text variables and corona.Rows correspond to each explanatory variable for which either the b3 or b4 coe�cient in(4) is signi�cant at the 10% level or better. EV column shows the impact of a one standarddeviation change in the explanatory variable in units of standard deviation of the VIXindex. EV*VIXl1 column shows the impact of a unit increase in V IX10 on the value ofEV. The V IXl1 column shows the e�ect of a one standard deviation increase in V IX10 inunits of standard deviation of the market variable. Numbers in parentheses are t-statitics,and the numbers in square brackets underneath the adjusted R2s are joint F-tests. Bothsets of tests use Newey-West standard errors with 3 lags. The R2 column shows adjustedR2s, and Chg R2 shows the di�erence in adjusted R2s between the speci�cation in (4)and a version of (4) which sets b4 = 0. The Break Date column shows the structuralbreak date of (4) using the procedure from Section 4.2, and the maximal Chow statistic isshown in square backets, with sigin�cance levels obtained from the Andrews (1993, 2003)distribution indicated via the *s. Signi�cance at the 10%, 5%, or 1% levels is indicatedvia *,**,*** respectively.
Contemp. changes in VIX index: early subperiod, VIX interact
const ht ht−1 EV EV*VIXl1 VIXl1 R2 Chg R2 Break Date
sent 42.398*** -0.768*** -0.206 -0.701*** -0.081*** -5.227*** 0.579 0.171 2020-03-16(2.84) (-5.43) (-1.52) (-5.22) (-4.82) (-4.42) [0.000] [54.29***]
sent_sd 84.147** -0.825*** -0.245 0.857*** 0.117*** -9.662*** 0.326 0.089 2020-03-16(2.23) (-2.59) (-1.23) (2.64) (3.57) (-3.62) [0.001] [28.04***]
s_sports 7.050* -0.810*** -0.438** -1.041*** -0.025 -1.003*** 0.368 0.014 2020-03-16(1.69) (-3.33) (-2.18) (-3.23) (-1.59) (-3.30) [0.001] [20.25**]
s_markets 0.772 -0.737*** -0.357* -0.798*** -0.069*** -0.892** 0.393 0.093 2020-03-16(0.09) (-3.44) (-1.66) (-3.56) (-4.93) (-2.08) [0.000] [24.65***]
s_health 8.645** -0.752*** -0.451** -1.099*** -0.113*** -1.181*** 0.410 0.080 2020-03-16(2.18) (-3.20) (-2.26) (-3.38) (-3.91) (-3.45) [0.016] [43.62***]
s_europe 14.929 -0.563*** -0.347 -0.829*** -0.064 -2.205 0.397 0.037 2020-03-11(0.82) (-3.64) (-1.62) (-4.54) (-1.56) (-1.45) [0.000] [17.66*]
s_oil & comm 50.372*** -0.435*** -0.167 -0.460*** -0.187*** -4.011*** 0.526 0.325 2020-03-13(3.53) (-2.60) (-0.94) (-2.63) (-4.05) (-3.77) [0.000] [36.76***]
s_currency -5.863 -0.621*** -0.216 -0.681** -0.038 -0.204 0.301 -0.000 2020-03-17(-0.83) (-3.57) (-0.93) (-2.15) (-0.96) (-0.32) [0.006] [16.74*]
corona 10.840*** -0.432*** -0.153 0.163* 0.079*** -0.769*** 0.476 0.276 2020-03-16(3.11) (-3.32) (-1.60) (1.73) (6.79) (-2.89) [0.000] [40.40***]
121C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 8
Contemporaneous regression of daily changes in US high-yield index on text variablesand corona. Rows correspond to each explanatory variable for which either the b3 or b4coe�cient in (4) is signi�cant at the 10% level or better. EV column shows the impact of aone standard deviation change in the explanatory variable in units of standard deviation ofthe US high-yield index. EV*VIXl1 column shows the impact of a unit increase in V IX10
on the value of EV. The V IXl1 column shows the e�ect of a one standard deviationincrease in V IX10 in units of standard deviation of the market variable. Numbers inparentheses are t-statitics, and the numbers in square brackets underneath the adjustedR2s are joint F-tests. Both sets of tests use Newey-West standard errors with 3 lags.The R2 column shows adjusted R2s, and Chg R2 shows the di�erence in adjusted R2sbetween the speci�cation in (4) and a version of (4) which sets b4 = 0. The Break Datecolumn shows the structural break date of (4) using the procedure from Section 4.2, andthe maximal Chow statistic is shown in square backets, with sigin�cance levels obtainedfrom the Andrews (1993, 2003) distribution indicated via the *s. Signi�cance at the 10%,5%, or 1% levels is indicated via *,**,*** respectively.
Contemp. changes in US high-yield index: early subperiod, VIX interact
const ht ht−1 EV EV*VIXl1 VIXl1 R2 Chg R2 Break Date
sent -9.038*** -0.184 -0.002 0.603*** 0.104*** 6.551*** 0.542 0.272 2020-03-04(-5.49) (-1.13) (-0.03) (10.65) (7.38) (6.94) [0.000] [41.69***]
sent_sd -19.528** -0.669 0.063 -1.100** -0.180*** 14.641** 0.329 0.162 2020-03-19(-2.57) (-1.47) (0.51) (-2.49) (-2.59) (2.57) [0.011] [35.46***]
s_sports -0.646 -0.299 -0.154** 0.645** 0.035* 0.663 0.272 0.049 2020-03-23(-0.79) (-1.38) (-2.08) (2.46) (1.66) (1.24) [0.001] [14.86]
s_markets -0.555 -0.374* -0.048 1.079*** 0.110*** 1.515*** 0.538 0.254 2020-03-23(-0.66) (-1.72) (-0.42) (5.49) (5.45) (3.26) [0.000] [31.05***]
s_health -2.147* -0.313 -0.117 1.245*** 0.171*** 1.812** 0.493 0.194 2020-03-19(-1.84) (-1.54) (-1.19) (3.80) (3.18) (2.40) [0.000] [25.82***]
s_europe -3.948* -0.146 -0.436*** 0.676*** 0.101*** 2.995** 0.336 0.122 2020-03-23(-1.87) (-0.64) (-5.68) (4.40) (2.99) (2.20) [0.000] [27.21***]
s_oil & comm -6.551*** 0.060 0.126 1.073*** 0.223*** 4.565*** 0.659 0.490 2020-03-23(-5.93) (0.55) (1.43) (6.36) (6.91) (7.01) [0.000] [39.53***]
s_corp & govt US -0.741 -0.171 -0.194* 0.338** 0.031 0.743 0.178 -0.003 2020-03-23(-0.46) (-0.60) (-1.67) (2.01) (1.26) (0.72) [0.040] [22.63**]
corona -0.697 -0.089 -0.219*** -0.222*** -0.076*** 0.370 0.418 0.251 2020-03-23(-1.45) (-0.49) (-2.86) (-2.99) (-7.47) (1.17) [0.000] [28.31***]
122C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 9
Contemporaneous regression of daily changes in 2-year Treasury yield on text variablesand corona. Rows correspond to each explanatory variable for which either the b3 or b4coe�cient in (4) is signi�cant at the 10% level or better. EV column shows the impact ofa one standard deviation change in the explanatory variable in units of standard deviationof the 2-year Treasury yield. EV*VIXl1 column shows the impact of a unit increase inV IX10 on the value of EV. The V IXl1 column shows the e�ect of a one standard deviationincrease in V IX10 in units of standard deviation of the market variable. Numbers inparentheses are t-statitics, and the numbers in square brackets underneath the adjustedR2s are joint F-tests. Both sets of tests use Newey-West standard errors with 3 lags.The R2 column shows adjusted R2s, and Chg R2 shows the di�erence in adjusted R2sbetween the speci�cation in (4) and a version of (4) which sets b4 = 0. The Break Datecolumn shows the structural break date of (4) using the procedure from Section 4.2, andthe maximal Chow statistic is shown in square backets, with sigin�cance levels obtainedfrom the Andrews (1993, 2003) distribution indicated via the *s. Signi�cance at the 10%,5%, or 1% levels is indicated via *,**,*** respectively.
Contemp. changes in 2-year Treasury yield: early subperiod, VIX interact
const ht ht−1 EV EV*VIXl1 VIXl1 R2 Chg R2 Break Date
sent -0.597* -0.050 0.587*** 0.616*** 0.093** 6.439** 0.229 0.125 2020-03-09(-1.76) (-0.64) (2.74) (3.65) (2.18) (2.14) [0.003] [37.47***]
s_sports -0.045 -0.167 0.231* 1.129** -0.006 0.820 0.148 -0.023 2020-03-09(-0.64) (-1.32) (1.78) (2.53) (-0.28) (1.02) [0.000] [29.60***]
s_central bank 0.046 -0.112 0.174 0.615* -0.029 -0.211 0.046 -0.007 2020-03-09(0.53) (-0.71) (0.88) (1.66) (-0.93) (-0.25) [0.236] [37.22***]
s_markets -0.187** -0.141 0.278 0.426* 0.078*** 1.714** 0.062 0.094 2020-03-09(-2.48) (-0.87) (1.55) (1.82) (2.90) (2.57) [0.000] [36.28***]
s_oil & comm -0.388*** 0.087 0.375*** 0.817*** 0.169*** 3.776*** 0.298 0.328 2020-03-09(-4.53) (0.81) (2.69) (4.45) (5.37) (5.05) [0.000] [44.46***]
s_currency -0.187*** -0.355*** 0.296*** 1.262*** 0.148*** 2.627*** 0.458 0.409 2020-03-12(-2.79) (-2.58) (2.67) (5.64) (4.69) (3.87) [0.000] [44.29***]
s_credit -0.169* 0.113 0.457** 0.576* 0.061* 1.784* 0.083 0.063 2020-03-09(-1.94) (1.28) (2.04) (1.94) (1.78) (1.87) [0.326] [45.22***]
s_corp future -0.128 -0.171 0.166 0.531** 0.089*** 1.292** 0.109 0.124 2020-03-09(-1.46) (-1.01) (0.90) (2.46) (2.72) (2.17) [0.045] [40.15***]
123C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 10
Contemporaneous regression of daily changes in 10-year Treasury yield on text variablesand corona. Rows correspond to each explanatory variable for which either the b3 or b4coe�cient in (4) is signi�cant at the 10% level or better. EV column shows the impact ofa one standard deviation change in the explanatory variable in units of standard deviationof the 10-year Treasury yield. EV*VIXl1 column shows the impact of a unit increase inV IX10 on the value of EV. The V IXl1 column shows the e�ect of a one standard deviationincrease in V IX10 in units of standard deviation of the market variable. Numbers inparentheses are t-statitics, and the numbers in square brackets underneath the adjustedR2s are joint F-tests. Both sets of tests use Newey-West standard errors with 3 lags.The R2 column shows adjusted R2s, and Chg R2 shows the di�erence in adjusted R2sbetween the speci�cation in (4) and a version of (4) which sets b4 = 0. The Break Datecolumn shows the structural break date of (4) using the procedure from Section 4.2, andthe maximal Chow statistic is shown in square backets, with sigin�cance levels obtainedfrom the Andrews (1993, 2003) distribution indicated via the *s. Signi�cance at the 10%,5%, or 1% levels is indicated via *,**,*** respectively.
Contemp. changes in 10-year Treasury yield: early subperiod, VIX interact
const ht ht−1 EV EV*VIXl1 VIXl1 R2 Chg R2 Break Date
sent -0.850 -0.060 0.310* 0.574*** 0.092* 6.493* 0.137 0.111 2020-03-17(-1.44) (-0.39) (1.95) (3.04) (1.66) (1.67) [0.006] [22.38**]
s_sports -0.043 -0.176 -0.054 0.891* -0.015 0.581 0.048 -0.017 2020-03-18(-0.38) (-1.35) (-0.38) (1.70) (-0.58) (0.61) [0.185] [26.52***]
s_markets -0.393*** -0.408*** -0.016 0.960*** 0.139*** 3.143*** 0.277 0.357 2020-03-17(-3.47) (-3.07) (-0.12) (3.90) (4.52) (4.03) [0.000] [23.82***]
s_health -0.312 -0.178* 0.070 0.940* 0.141 2.234 0.092 0.114 2020-03-17(-1.48) (-1.87) (0.70) (1.83) (1.48) (1.48) [0.075] [31.37***]
s_oil & comm -0.742*** 0.000 0.181* 1.110*** 0.232*** 5.292*** 0.543 0.623 2020-03-13(-6.25) (0.00) (1.85) (6.64) (8.30) (7.43) [0.000] [22.33**]
s_currency -0.286*** -0.279** 0.110 1.127*** 0.139*** 2.693*** 0.383 0.404 2020-03-17(-3.64) (-2.33) (0.79) (5.50) (4.43) (4.47) [0.000] [36.30***]
s_corp & govt US 0.422** -0.171 -0.201 0.153 -0.077** -2.522* 0.054 0.074 2020-03-20(2.07) (-1.15) (-1.09) (0.74) (-2.21) (-1.85) [0.005] [28.99***]
s_corp future -0.258** -0.211 -0.092 0.457*** 0.095*** 1.661*** 0.103 0.184 2020-03-17(-2.45) (-1.42) (-0.45) (3.02) (3.67) (3.45) [0.000] [18.95**]
124C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 11
Test of whether the variable in a given row Granger causes the variables in the columns of the table. The table shows c3(EV) and c4 (EV*VIX) from the speci�cations in (7). The c3 coe�cient is normalized to show standard deviation changesin the response variable due to a single standard deviation change in the test variable. c4 shows changes in c3 due to a unitchange in the VIX. T-statistics are shown in parentheses. The d_[market] variables refer to daily returns or changes inthe particular market series. Standard errors use Newey-West with three lags. Signi�cance at the 10%, 5%, or 1% levels isindicated via *,**,*** respectively. Data are daily.
Granger causality tests: Market changes to text measures in early subsample
sent sent_sd s_sports s_central bank s_markets s_health s_europe
d_sp500
EV 0.882*** -0.115 0.292** 0.150 0.515*** 0.344*** 0.531***(EV t) (5.22) (-1.04) (2.53) (1.49) (2.84) (2.86) (2.74)EV*VIX -0.044*** -0.001 -0.010 -0.015*** -0.018* -0.017** -0.035***(EV*VIX t) (-4.76) (-0.22) (-1.21) (-4.00) (-1.80) (-2.54) (-3.60)
d_vix
EV -0.788*** 0.196* -0.225** -0.091 -0.537*** -0.256** -0.470***(EV t) (-5.68) (1.73) (-1.98) (-1.20) (-3.16) (-2.44) (-3.25)EV*VIX 0.038*** -0.003 0.007 0.013*** 0.019** 0.012** 0.030***(EV*VIX t) (5.36) (-0.47) (1.09) (5.53) (2.22) (2.30) (4.27)
d_hy
EV 0.803*** -0.073 0.169 0.026 0.661*** 0.354** 0.532**(EV t) (3.14) (-0.52) (0.82) (0.29) (2.87) (2.27) (2.48)EV*VIX -0.041*** -0.003 -0.009 -0.010** -0.024** -0.017** -0.030***(EV*VIX t) (-3.37) (-0.38) (-0.94) (-2.28) (-2.17) (-2.44) (-2.83)
d_gt2
EV 0.627*** -0.013 0.145** 0.140** 0.344** 0.281*** 0.357***(EV t) (4.98) (-0.14) (2.00) (1.99) (2.53) (4.36) (3.28)EV*VIX -0.039*** -0.008 -0.003 -0.010** -0.022** -0.018*** -0.032***(EV*VIX t) (-4.76) (-1.32) (-0.53) (-2.00) (-2.53) (-4.40) (-4.30)
d_gt10
EV 0.854*** 0.016 0.153* 0.134 0.572*** 0.429*** 0.436***(EV t) (5.32) (0.15) (1.81) (1.45) (3.90) (6.22) (2.71)EV*VIX -0.049*** -0.009 -0.006 -0.010** -0.031*** -0.025*** -0.032***(EV*VIX t) (-5.43) (-1.41) (-0.99) (-2.19) (-4.06) (-6.78) (-3.61)
125C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 11
Test of whether the variable in a given row Granger causes the variables in the columns of the table. The table shows c3(EV) and c4 (EV*VIX) from the speci�cations in (7). The c3 coe�cient is normalized to show standard deviation changesin the response variable due to a single standard deviation change in the test variable. c4 shows changes in c3 due to a unitchange in the VIX. T-statistics are shown in parentheses. The d_[market] variables refer to daily returns or changes inthe particular market series. Standard errors use Newey-West with three lags. Signi�cance at the 10%, 5%, or 1% levels isindicated via *,**,*** respectively. Data are daily.
Granger causality tests: Market changes to text measures in early subsample
s_oil & comm s_currency s_credit s_corp & govt US s_corp actual s_corp future s_credit1 corona
d_sp500
EV 0.013 0.282*** 0.126 0.279** 0.090 0.086 -0.085 -0.017(EV t) (0.19) (2.60) (1.01) (2.30) (0.54) (0.55) (-0.37) (-0.11)EV*VIX -0.001 -0.013** -0.009 -0.018** -0.007 -0.002 0.016 0.007(EV*VIX t) (-0.14) (-2.32) (-1.52) (-2.56) (-0.87) (-0.27) (1.37) (0.70)
d_vix
EV -0.098 -0.342*** -0.165 -0.256** -0.049 -0.110 -0.029 -0.000(EV t) (-1.61) (-3.67) (-1.58) (-2.03) (-0.42) (-0.64) (-0.15) (-0.00)EV*VIX 0.004 0.015*** 0.010** 0.016** 0.005 0.003 -0.011 -0.010(EV*VIX t) (1.02) (3.69) (2.10) (2.37) (0.94) (0.37) (-1.14) (-1.17)
d_hy
EV 0.148 0.520** -0.083 0.132 -0.199 -0.007 -0.174 0.259(EV t) (0.94) (2.49) (-0.54) (0.67) (-0.97) (-0.04) (-0.66) (0.98)EV*VIX -0.011 -0.024*** -0.003 -0.013 0.007 0.001 0.011 -0.008(EV*VIX t) (-1.20) (-2.63) (-0.33) (-1.34) (0.65) (0.07) (0.82) (-0.51)
d_gt2
EV -0.008 0.330*** 0.046 0.269*** -0.052 0.022 0.054 0.252(EV t) (-0.08) (2.72) (0.54) (3.95) (-0.45) (0.24) (0.54) (1.55)EV*VIX -0.003 -0.017** -0.009* -0.013** -0.005 0.001 -0.001 -0.008(EV*VIX t) (-0.54) (-2.43) (-1.70) (-2.44) (-0.71) (0.09) (-0.11) (-0.87)
d_gt10
EV 0.078 0.507*** 0.116 0.230* -0.142 0.039 0.048 0.187(EV t) (0.75) (4.09) (1.18) (1.87) (-0.92) (0.31) (0.34) (1.64)EV*VIX -0.007 -0.027*** -0.013** -0.013* 0.001 -0.001 -0.005 -0.003(EV*VIX t) (-1.19) (-4.25) (-2.56) (-1.72) (0.16) (-0.12) (-0.59) (-0.38)
126C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 12
Test of whether the variable in a given row Granger causes the variables in the columns ofthe table. The table shows c3 (EV) and c4 (EV*VIX) from the speci�cations in (7). Thec3 coe�cient is normalized to show standard deviation changes in the response variabledue to a single standard deviation change in the test variable. c4 shows changes in c3due to a unit change in the VIX. T-statistics are shown in parentheses. The d_[market]variables refer to daily returns or changes in the particular market series. Standard errorsuse Newey-West with three lags. Signi�cance at the 10%, 5%, or 1% levels is indicatedvia *,**,*** respectively. Data are daily.
Granger causality tests: Text measures to market changes in early subsample
d_sp500 d_vix d_hy d_gt2 d_gt10
sent
EV -0.129 -0.085 -0.347** -0.174 -0.218*(EV t) (-0.80) (-0.53) (-2.48) (-1.38) (-1.66)EV*VIX -0.083*** 0.058** -0.078*** -0.041** -0.057***(EV*VIX t) (-2.90) (2.24) (-3.08) (-2.38) (-2.76)
sent_sd
EV 1.137*** -1.098*** 1.181*** 0.906*** 1.231***(EV t) (6.31) (-5.57) (3.07) (3.25) (3.01)EV*VIX 0.175*** -0.169*** 0.155*** 0.135*** 0.174***(EV*VIX t) (5.50) (-5.49) (2.82) (3.45) (3.27)
s_sports
EV -0.613* 0.355 -1.333*** 0.068 -0.364(EV t) (-1.67) (0.87) (-3.20) (0.13) (-0.69)EV*VIX -0.067** 0.055** -0.090*** -0.056** -0.075***(EV*VIX t) (-2.47) (2.55) (-4.48) (-2.19) (-2.81)
s_central bank
EV 0.096 -0.169 -0.481 0.384 0.284(EV t) (0.19) (-0.40) (-1.49) (0.89) (0.84)EV*VIX -0.003 -0.003 -0.002 -0.020 -0.039*(EV*VIX t) (-0.15) (-0.15) (-0.09) (-0.94) (-1.86)
s_markets
EV -0.253 0.174 -0.442* -0.892*** -1.077**(EV t) (-1.06) (0.97) (-1.78) (-3.54) (-2.56)EV*VIX -0.061*** 0.054*** -0.078*** -0.108*** -0.133**(EV*VIX t) (-3.08) (3.48) (-2.76) (-3.31) (-2.44)
s_health
EV -0.163 0.111 -0.892*** -0.714 -0.834(EV t) (-0.65) (0.65) (-3.93) (-1.43) (-1.29)EV*VIX -0.097** 0.088*** -0.176*** -0.100 -0.140(EV*VIX t) (-2.06) (2.71) (-3.80) (-1.20) (-1.31)
s_europe
EV 0.136 -0.306 0.012 -0.295 -0.617**(EV t) (0.29) (-0.74) (0.04) (-1.35) (-2.14)EV*VIX -0.013 -0.000 0.000 -0.051 -0.114**(EV*VIX t) (-0.33) (-0.00) (0.01) (-1.60) (-2.06)
s_oil & commEV -0.509 0.345 0.019 -0.808*** -0.863***(EV t) (-1.18) (1.28) (0.04) (-3.20) (-2.84)
127C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 12
Test of whether the variable in a given row Granger causes the variables in the columns ofthe table. The table shows c3 (EV) and c4 (EV*VIX) from the speci�cations in (7). Thec3 coe�cient is normalized to show standard deviation changes in the response variabledue to a single standard deviation change in the test variable. c4 shows changes in c3due to a unit change in the VIX. T-statistics are shown in parentheses. The d_[market]variables refer to daily returns or changes in the particular market series. Standard errorsuse Newey-West with three lags. Signi�cance at the 10%, 5%, or 1% levels is indicatedvia *,**,*** respectively. Data are daily.
Granger causality tests: Text measures to market changes in early subsample
d_sp500 d_vix d_hy d_gt2 d_gt10
s_oil & commEV*VIX -0.050 0.019 0.022 -0.119*** -0.128**(EV*VIX t) (-0.52) (0.31) (0.22) (-2.85) (-2.42)
s_currency
EV 0.120 -0.190 -0.162 0.420 0.546(EV t) (0.36) (-0.62) (-0.49) (1.02) (1.52)EV*VIX -0.062 0.050 -0.046 0.046 0.068(EV*VIX t) (-1.46) (1.28) (-0.98) (0.96) (1.49)
s_credit
EV 0.035 -0.182 -0.006 0.085 0.025(EV t) (0.10) (-0.61) (-0.02) (0.26) (0.08)EV*VIX -0.014 -0.002 0.010 -0.011 -0.019(EV*VIX t) (-0.33) (-0.06) (0.28) (-0.26) (-0.43)
s_corp & govt US
EV 0.312 -0.320 0.379* -0.187 0.118(EV t) (1.15) (-1.09) (1.78) (-0.68) (0.34)EV*VIX 0.035 -0.039 0.051 -0.025 -0.018(EV*VIX t) (0.88) (-1.00) (1.61) (-0.86) (-0.60)
s_corp actual
EV 0.297** -0.403** 0.162 0.451 0.269(EV t) (2.51) (-2.52) (0.94) (1.64) (0.98)EV*VIX 0.013 -0.026 0.027 0.042 0.020(EV*VIX t) (0.64) (-1.30) (0.99) (1.03) (0.43)
s_corp future
EV 0.017 -0.166 0.040 0.308 0.507(EV t) (0.07) (-0.98) (0.19) (1.08) (1.64)EV*VIX -0.004 -0.001 0.005 0.066 0.104*(EV*VIX t) (-0.10) (-0.02) (0.12) (1.38) (1.91)
s_credit1
EV -0.066 0.093 -0.074 0.092 0.122(EV t) (-0.30) (0.44) (-0.38) (0.47) (0.52)EV*VIX -0.033 0.034 -0.013 -0.002 -0.009(EV*VIX t) (-1.43) (1.55) (-0.49) (-0.09) (-0.35)
corona
EV 0.830*** -0.782*** 0.209* 0.403** 0.519***(EV t) (3.09) (-2.88) (1.73) (2.17) (3.25)EV*VIX -0.007 0.011 0.026 0.032* 0.069***(EV*VIX t) (-0.33) (0.64) (1.30) (1.86) (3.22)
128C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
68-
128
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Covid Economics Issue 38, 16 July 2020
Copyright: Graziella Bertocchi and Arcangelo Dimico
COVID-19, race, and redlining1
Graziella Bertocchi2 and Arcangelo Dimico3
Date submitted: 7 July 2020; Date accepted: 9 July 2020
Using individual, race-disaggregated, and georeferenced death data collected by the Cook County Medical Examiner, we look at the impact of COVID-19 on African Americans and at its determinants. First, we provide evidence that - as of June 16, 2020 - blacks in Cook County are dying from COVID-19 at a rate 1.3 times higher than their population share. Second, by combining the spatial distribution of mortality with the redlining maps for the Chicago area, we establish that - after the epidemic outbreak - historically lower-graded neighborhoods display a sharper increase in mortality, driven by blacks. Thus, we uncover a persistence influence of the racial segregation induced by the lending practices of the 1930s, by way of a diminished resilience of African Americans to the COVID-19 shock. Such influence is channeled through socioeconomic status and household composition, and magnified in combination with a higher black share.
1 We would like to thank Chris Colvin, Guido Friebel, Marianne Wanamaker, and seminar participants at Queen's University Belfast and Goethe University for helpful comments.
2 Professor of Economics, University of Modena and Reggio Emilia, EIEF, CEPR, CHILD, Dondena, GLO, and IZA.
3 Senior Lecturer in Economics, Queen's University Belfast, GLO, IZA, CHaRMS, and QUCEH.
129C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
1 Introduction
Ever since when The Atlantic magazine (Kendi, 2020), on March 4, 2020, first launched
a cry for attention to the disproportionate impact of COVID-19 on African Americans,
the issue has taken center stage in the debate on the socioeconomic implications of the
pandemic. Nevertheless, a race-disaggregated analysis of individual data has so far been
lacking due to data unavailability.
The urgency of the racial issue and the need for data collection to account for ethnic
and racial factors has been widely recognized also within the medical and epidemiological
literature. Preliminary evaluations suggest that the high risk of COVID-19 death for mi-
nority ethnic groups can be explained by pre-existing health conditions, such as diabetes,
obesity, hypertension, and asthma, that are more common among these groups, possibly
because of genetic and biological factors. However, the emerging consensus is that the
race differential in the prevalence of COVID-19 is also associated with socioeconomic cor-
relates. As argued by Yancy (2020), a large share of the black population in the US lives
in poor areas characterized by high unemployment, low housing quality, and unhealthy
living conditions, making low socioeconomic status a critical risk factor. Furthermore,
the higher prevalence of comorbidities in blacks is also itself associated with highly per-
sistent socioeconomic factors. Other relevant health-related and behavioral risk factors,
such as smoking, drinking, and drug abuse, are deeply ingrained in cultural norms that
are also driven by social inequalities. African Americans suffer for further disadvantage
in their ability to adhere to social distancing norms, as working from home, avoiding
public transportation, and finding refuge in second homes away from crowded cities, are
indeed privileges that are denied to the majority of them. Anecdotal evidence also sug-
gests the possibility of a different response by health practitioners to black individuals
showing COVID-19 symptoms. All the above considerations, as summarized by Yancy
(2020) resonate with the long-standing debate on the racial disparities that are deeply
entrenched in US history and point to a need to account for race in COVID-19 research
and to investigate the role not only of biological factors but also of socioeconomic ones,
while acknowledging that the latter may be rooted on inequalities that have been long
neglected.
In this paper, we take advantage of an unexplored and extraordinarily detailed source
of information on daily deaths from COVID-19 that includes race among a wide array
of individual characteristics. The data are collected by the Medical Examiner’s Officer
and made available by the Government of Cook County, Illinois, the county that hosts—
among others—the City of Chicago.1
1Since April 15, after some states started to report data, the COVID Tracking Project at The Atlantic
130C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
On the basis of the Cook County data we can provide, to start with, evidence of how
race affects COVID-19 individual outcomes. Furthermore, we can illustrate the correlates
and dig into the historical determinants of such outcomes. The higher COVID-19 death
toll paid by black Americans has been linked to the “redlining”policies introduced by the
Home Owners Loan Corporation (HOLC) in the 1930s.2 These policies are believed to
have favored the development of segregated neighborhoods plagued by unemployment,
low housing quality, and unhealthy living conditions. In a public speech following the
death of George Floyd, in the midst of the pandemic, former President Obama has also
attributed the explosion of protests to the history of racial discrimination, including
redlining.3 By combining the information on COVID-19 fatalities in Cook County with
the redlining maps for the Chicago area, we can assess the explanatory power of the
induced vulnerabilities that are rooted in history.
In more detail, the first contribution of the paper is to document how race affects
COVID-19 mortality in Cook County, on the basis of individual data that also account for
age, gender, and pre-existing health conditions. The analysis confirms that, as reported
by media and public authorities, blacks are overrepresented in terms of COVID-19 related
deaths since—as of June 16, 2020—they constitute 35 percent of the deaths against a
black population share of only 27 percent in our sample. In other words, blacks are dying
at a rate 1.3 times higher than their population share. This figure is lower than those
previously reported at a early stage, because blacks were the first to be hit by the virus,
while other groups followed up only later on.
Second, by exploiting the fact that the Cook County Medical Examiner also pro-
vides the geographical coordinates of the home address of each individual that died from
COVID-19, we combine the spatial distribution of mortality with the redlining maps for
the Chicago area, to assess whether the higher vulnerability of blacks to the disease can
indeed be found in socioeconomic inequalities rooted in history, rather than in biological
conditions. Using cross-sectional information about individual deaths from COVID-19,
we show that the probability that an individual who died from COVID-19 is black is much
larger in lower graded areas, that is, those that were historically redlined and yellowlined,
even after controlling for demographic and socioeconomic factors as well as pre-existing
has been providing updated state-level race and ethnicity data on cases and deaths, with informationlimited to race and ethnicity. See https://covidtracking.com/race. On July 5, 2020 The New YorkTimes (Oppel et al., 2020) reported data on 640,000 individual cases, by race, ethnicity and homecounty, collected through May 28. The data were acquired after filing a Freedom of Information Act suitagainst the Centers for Disease Control and Prevention.
2See Eligon et al. (2020) in The New York Times.3“They are the outcomes not just of the immediate moments in time, but they are the result of a long
history of slavery and Jim Crow and redlining and institutionalized racism that too often have been theplague, the original sin, of our society.” See https://www.youtube.com/watch?v=q qB6SsErpA.
131C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
conditions. However, where we control for the black population share, it fully absorbs the
influence of the HOLC policies, suggesting that the latter did induce segregation along
racial lines. In other words, the black share fully captures the effect of residence in a
low-graded area. A warning is in order since the cross-sectional analysis, being based
solely on information about those that died from COVID-19, is severely biased because
of sample selection. This limitation motivates our next set of results.
Third, to establish a causal link between the vulnerabilities induced by historical
redlining and the racial gaps in COVID-19 outcomes, and to disentangle socioeconomic
factors from other determinants which normally correlate with African Americans, we rely
on an event study analysis based on a weekly balanced panel of deaths from any cause
that we assemble at census block group level over the period from January 1, 2020 to June
16, 2020. In this way, we can identify the impact of the asymmetric shock introduced
by COVID-19 on lower-grade block groups—the treated group—after treatment initia-
tion, that is, after the epidemic outbreak. Furthermore, we can test for pre-treatment
differences while controlling for block group fixed effects which should filter out the ef-
fect of invariant socioeconomic factors. Consistent with the cross-sectional evidence, to
determine the treatment we refer to block groups that belong to HOLC areas graded
either C or D, that is, either yellow or redlined. We find that, while no pre-treatment
differences are detected, mortality in treated neighborhoods increases sharply after the
epidemic shock, and the increase is driven by the death toll of African Americans. The
implied magnitude of the estimated effect is an increase in the number of black deaths
by almost 60 percent.
We also carry out a heterogeneity analysis focused on the degree of vulnerability to
shocks, that we measure using data provided by the Centers for Disease Control and Pre-
vention (CDC). The analysis reveals that the channels through which historical redlining
influences COVID-19 outcomes are socioeconomic characteristics, in particular personal
income and the population share below poverty, and household composition, in particular
the population shares of the elderly and of single parent households. Strikingly, however,
the influence of these factors manifests itself only in combination with a higher black
share. We complete our investigation with a battery of robustness checks, extensions,
and falsification tests.
Overall, the evidence we collect points to a persistence influence of the racial segre-
gation introduced by the discriminatory lending practices of the 1930s, that operates by
way of an asymmetric effect of the epidemic shock, which is in turn channeled through
a diminished resilience of African Americans. Far from being determined by genetic and
biological factors, their vulnerability can be linked to socioeconomic status and household
composition, as channels through which the legacy of the past manifests itself.
132C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
The paper is organized as follows. Section 2 presents related literature. Section 3
summarizes background information on historical redlining policies and Cook County.
Section 4 describes the data. Section 5 presents cross-sectional evidence. Section 6 intro-
duces an event study approach and the corresponding baseline results, while Section 7 is
devoted to robustness checks and Section 8 to the heterogeneity analysis. In Section 9 we
extend the investigation to the Hispanic population. Section 10 presents two falsification
tests respectively based on deaths in the years 2017 to 2019 and Spanish flu deaths in
1918. Section 11 concludes. The Appendix contains additional figures and tables.
2 Related literature
We contribute to several streams of the literature. The first is the literature on the racial
impact of COVID-19. The so far largest epidemiological study on the racial impact of
COVID-19 has been performed in the UK based on the medical records of more than
17 million individuals (OpenSAFELY Collaborative, 2020, working on behalf of NHS
England).4 The study confirms that pre-existing medical conditions such as diabetes or
deprivation are linked to a higher likelihood of in-hospital death, but also that clinical risk
factors alone cannot explain the observed disparities. Again with a focus on the UK, Bhala
et al. (2020) also support structural, socioeconomic and environmental explanations,
rather than biological ones, of the racial differences in COVID-19 susceptibility. Attention
is called to the fact that ethnic/racial minorities hold highly-exposed jobs in health and
social care, retail, and public transport, and to cultural habits including the approach to
worship and the multigenerational family structure. On the basis of medical data from
COVID-19 patients at a hospital in Louisiana, Price-Haywood et al. (2020) find that
blacks were overrepresented among patients and fatalities, but did not show higher in-
hospital mortality than whites after controlling for differences in sociodemographic (i.e.,
insurance type and zip code of residence) and clinical characteristics on admission.
Despite the above mentioned growing body of contributions within medicine and epi-
demiology, within the economics field the literature on the racial impact of COVID-19
is still limited. Borjas (2020), Schmitt-Grohe et al. (2020), and Almagro and Orane-
Hutchinson (2020) account for the racial dimension while looking at the demographic and
socioeconomic correlates of the COVID-19 epidemics, with a focus on testing incidence
and infections, but not deaths, across New York City zip codes. Across US counties,
Desmet and Wacziarg (2020) find a positive correlation between the shares of African
Americans and Hispanics and both the number of cases and the number of deaths, a
4In the UK, the racial issue came to public attention when the first eleven doctors who died fromCOVID-19 were all reported to belong to black, Asian, and minority ethnic communities (Kirby, 2020).
133C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
finding which is confirmed by McLaren (2020) for African Americans’ deaths even after
controlling for education, occupation, and commuting patterns. Using the 2017 wave of
the Panel Study of Income Dynamics to examine the prevalence of specific health con-
ditions, Wiemers et al. (2020) show evidence of large disparities across race-ethnicity
and socioeconomic status in the prevalence of conditions which are associated with the
risk of severe complications from COVID-19. Using CPS microdata on unemployment,
Couch et al. (2020) show that African Americans were only slightly disproportionately
impacted by COVID-19, while Latin Americans were hardly hit. With reference to the
UK, Platt and Warwick (2020), Sa (2020), and White and Nafilyan (2020) report de-
scriptive evidence on vulnerability factors, infections, and deaths, for ethnic minorities
including Black Caribbean and Black African.5
Our second contribution is to research on the long-term influence of redlining policies,
which have been the object of investigation not only within the field of economics but
also within medicine, history, and law. Zenou and Boccard (2000), Appel and Nickerson
(2016), Krimmel (2018), Aaronson et al. (2017), Mitchell and Franco (2018), and Anders
(2019) respectively look at the effects on unemployment, home prices, homeownership,
racial segregation, inequality, and crime. Krieger et al. (2020a, 2020b) and Nardone et
al. (2020) respectively associate historical redlining with higher risk of cancer, preterm
birth, and asthma, suggesting that this discriminatory practice might be contributing
to racial and ethnic disparities.6 Jackson (1980, 1985), Hillier (2003, 2005), and Greer
(2012, 2014) provide historical accounts of the activities of the HOLC and its influence
on American cities. Schill and Wachter (1995) and Nier (1999) offer a legal interpretation
of redlining and the induced spatial bias of federal housing law and policy.7
A third research stream that is relevant to our approach has looked at other pan-
demics. The long-term determinants of the HIV/AIDS pandemic, with special attention
to women, who represent within Africa the most vulnerable group, have been studied
by Anderson (2018), who links the higher female HIV rates to the tradition of com-
mon law, while Bertocchi and Dimico (2019), Loper (2019), and Cage and Rueda (2020)
respectively refer to the slave trade and the associated diffusion of polygyny, the prac-
tice of matrilinearity in ancestral societies, and the influence of the Christian missions
5For the much broader literature on racial discrimination in the US, see Lang and Kahn-Lang Spitzer(2020) for a survey and Nunn (2008), Bertocchi and Dimico (2012, 2014, 2020a), Bertocchi (2016), andreferences therein, on the legacy of slavery as its long-term determinant.
6On residential segregation and cancer see also Landrine et al. (2017). An extensive literatureconfronts racial and ethnic disparities in health care and health outcomes. See Institute of Medicine(2003) and Orsi et al. (2010).
7On the history of the HOLC, see Fishback et al. (2013). On the broader determinants, other thanredlining, of race segregation in the metropolitan areas of the US, see Boustan (2011) and Rothstein(2017). On the influence of zoning policies, with a focus on 1923 Chicago, see Shertzer et al. (2016).
134C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
established during the colonial period. The long-term economic, social and cultural con-
sequences of the “Spanish”influenza have been studied by Almond (2006), Karlsson et
al. (2014), Lin and Liu (2014), Helgertz and Bengtsson (2019), Aassve et al. (2020), and
Guimbeau et al. (2020), while Richardson and McBride (2009), Voitglander and Voth
(2012), Jedwab et al. (2016, 2019), and Alfani and Murphy (2017) have studied the Black
Death.
3 Historical background
3.1 Redlining
The Home Owners Loan Corporation (HOLC) was created in June 1933 by the US
Congress, in the aftermath of the Great Depression and within the first 100 days of the
Roosevelt administration, as part of a key package of New Deal policies aimed at rescuing
the housing and banking sectors through actions on the mortgage lending market. In the
general effort to revive the economy, housing policies were viewed as critical and were
therefore assigned a major role. The task of the HOLC was to refinance mortgages in
default to prevent foreclosures, as a response to the banking sector turmoil and the drastic
fall in home loans and ownership (Harriss, 1951). In 1934 the National Housing Act
established the Federal Housing Administration (FHA) to reinforce previous measures
and boost the market for single-family homes. With the goal of improving the accuracy
of real-estate appraisal and in turn standardizing the process of mortgage lending, credit
worthiness assessment, and mortgage support assignment, in 1935 the HOLC was asked
to create “Residential Security Maps”of 239 cities that ranked areas on the basis of default
risk. The ranking encompassed four levels. The safest areas, mostly consisting of newly-
build suburban neighborhoods, were labelled as “Best”, assigned to Type A, and outlined
in color green. “Still Desirable”areas were assigned to Type B and outlined in blue. The
next two levels included “Definitely Declining”areas, assigned to Type C and outlined in
yellow, and “Hazardous”areas assigned to Type D and outlined in red. Because of the
color used to highlight to the worst-assessed neighborhoods, those that ended up being
de facto denied any mortgage financing, the process came to be known as “redlining”.
The HOLC rankings were based on meticulous assessments and recording of neighbor-
hood characteristics including population growth, class and occupation of the inhabitants,
and block-by-block quality of the buildings (type, size, construction material, age, need
for repair, occupancy rate, owner-occupancy rate, past and predicted property prices,
rents, and sales and rental demand trends). Notably, the share of foreign and black fam-
ilies and the degree to which the neighborhood was deemed to be “infiltrated”were also
135C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
accurately recorded. Figure A1 in the Appendix reports the area description drafted by
the HOLC in 1939 for a D-rated neighborhood in South Chicago, summarily described
as “A 100 per cent negro development. (...) A blighted section”.
To assemble the mapping, the HOLC trained thousands of home appraisers and, in
the process, set standards for the development of a new approach to mortgage lending,
which were adopted and further refined in the Underwriting Manuals compiled by the
FHA (1938). The task for which the HOLC was created was undoubtedly fulfilled, as
the agency had a major impact on the subsequent expansion of real estate investments.
As documented by Harriss (1951), between 1933 and 1935 the HOLC received almost
1.9 million applications for home mortgage refinancing.8 Out of the 54 percent of them
that were accepted, the majority involved one- or two-family homes of modest size and
value and borrowers of relatively limited income. In the New York region, 44 percent of
the properties whose purchase was supported with a loan were located in neighborhoods
described as “native white”and 42 percent in “native white and foreign.”The fact that
only 1 percent of the applications covered properties in neighborhoods described as “na-
tive Negro”is attributed by Harriss to the low percentage of applicants from such areas,
which “doubtless reflects the fact that most Negroes (...) lived in rented quarters and did
not, therefore, fall within the limits of the HOLC programs.”
The direct and indeed intended consequences of redlining were to channel credit and
investment away from poorer areas and toward more affluent ones. As a result, the slums
deteriorated even farther. Over time, the practice is widely believed to have contributed to
the exacerbation and persistence of initial inequalities (Douglas Commission, 1968). After
the Second World War, racial segregation further intensified with the “white flight”from
the inner cities to the suburbs (Boustan, 2011). It was only with the Fair Housing Act of
1968, a provision of the Civil Rights Act, that housing segregation was outlawed, while
specific legislation to establish fair lending practices was only enacted in the 1970s with
the Equal Credit Opportunity Act (1974) and the Community Reinvestment Act (1977).
Throughout the process, the HOLC maps were deliberately hidden from public view,
even though they may have been shared selectively with realtors and lenders (Greer, 2012,
2014). The existence of the maps emerged later and became the subject of investigation
of the National Commission on Urban Problems (Douglas Commission, 1968), created
by President Johnson in 1965 “to study building codes, housing codes, zoning, local and
Federal tax policies and development standards” in order to “provide knowledge that would
be useful in dealing with slums, urban growth, sprawl and blight, and to insure decent
and durable housing.” But it was only much later that Jackson (1980, 1985), an urban
8Subsequently, the activities of the HOLC were devoted to loan management and repayments. Oper-ations were officially ceased in 1951 and termination was ordered in 1954.
136C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
historian, discovered the HOLC Residential Security Maps in the National Archives,
documenting what he describes as a system designed to apply “notions of ethnic and
racial worth to real-estate appraising on an unprecedented scale.” His discovery spurred
a renewed research effort aimed at identifying redlining as a key factor in perpetuating
racial disparities that are still observed up to the present day.
3.2 Cook County, Illinois
With over five million residents in 2019, Cook County is the most populous county in
Illinois and the second-most-populous county in the US after Los Angeles County, Cal-
ifornia. Over 40 percent of all residents of Illinois live in Cook County. The largest
of the county’s 135 municipalities is the City of Chicago—the third-most-populous US
city—followed by the City of Evanston. Overall, Cook County is highly urbanized and
densely populated. According to the United States Census Bureau,9 in 2019 non-Hispanic
whites, Hispanics, African Americans, and Asians were the most represented racial and
ethnic groups, respectively with 42.0, 25.6, 23.8, and 7.9 percent of the population,10 A
21.6 percent share of the population was under age 18 and 15.1 percent was above 65. A
13.8 fraction of the population, higher than the national average, was below the poverty
line. Health status disparities between black and white populations widened in Chicago
between 1990 and 2005 (Orsi et al., 2010). Politically, the county is heavily Democratic-
leaning, with a 73.9 share of the votes being cast for the Democratic Party in the 2016
presidential elections.
According to the Johns Hopkins University & Medicine Coronavirus Resource Cen-
ter,11 as of June 16, 2020 Cook County ranked third among US counties, after Kings and
Queens in New York, in terms of COVID-19 related deaths, with over 4,000 deaths, and
first in terms of confirmed cases, with over 80,000 cases. A strict stay-at-home order was
issued by Illinois Governor Pritzker on March 20 (effective March 21), four days after the
first COVID-19 related death, when the death toll was still limited to five.
Turning to the urban history of Cook County in the aftermath of the Great Depression,
in his 1933 dissertation on the evolution of land values in Chicago Hoyt (1933)—before
joining the FHA in 1934 as Principal Housing Economist—provided a ranking of races and
nationalities with respect to their beneficial effect upon land values. While acknowledg-
ing that such an effect may have been reflecting racial prejudice, he ranked Anglo-Saxons
and Northern Europeans at the top and Negroes, followed by Mexicans, at the bottom.
9See https://www.census.gov/quickfacts/fact/table/cookcountyillinois/PST120219.10Racial and ethnic data are based on self-identification. Reported figures are “alone”, that is, refer
to those individuals that self-classify themselves as belonging to a single racial or ethnic group.11See https://coronavirus.jhu.edu/.
137C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 1: Historical Maps of the Chicago AreaNote: The figure shows, on the left panel, the “Map of Chicago Showing Area Occupied by Predominant Racialor Nationality Groups, 1933”(Hoyt, 1933) and, on the right panel, the HOLC maps for the Chicago area (Nel-son et al., 2020), with green, blue, yellow, and red denoting respectively grade A, B, C, and D neighborhoods.
He also produced a map of Chicago (Figure 1, left panel) reporting the areas occupied
by predominant groups among the most recent immigrant waves, namely Negro, Italian,
Polish, Jewish, and Czechoslovakian newcomers. As the figure shows, blacks were con-
centrated in the so-called “Black Belt”on Chicago’s South Side, where they were forced
to settle from the beginning of the Great Migration, facing squalid housing conditions
and extremely high population densities (Greer, 2014).
By 1940, a large portion of Cook County was mapped by the HOLC. On the right
panel, Figure 1 shows the HOLC areas of the Chicago area as rendered in the Mapping
Inequality: Redlining in New Deal America 1935-1940 dataset by the University of Rich-
mond.12 As in the other American cities, the geography of redlining had a clear racial
connotation. The figure reveals that the same areas inhabited by blacks in the Hoyt map
were assigned the lowest grade and highlighted in red. Indeed those blocks were char-
acterized by houses lacking basic amenities such as access to water and heating (Greer,
2014). Krimmel (2018) estimates that, by 1940, 98 percent of the relatively small share
of blacks living in the city, about 8 percent, were redlined.
The programs that were implemented by the agency were welcome with wide support
from the Chicago white press, and were originally only weakly opposed by the black
12See Nelson et al. (2020) and https://dsl.richmond.edu/panorama/redlining/.
138C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
press, including the Chicago Defender. Nevertheless, the disinvestment induced by the
HOLC policies had consequences that extended well past the Second World War, with
white families easily obtaining mortgage insurance to move to the suburbs, and black ones
being relegated in blighted neighborhoods where the only financing opportunity consisted
in exploitative installment land contracts (Greer, 2012, 2014). As a result, as of 1967,
in the words of the Douglas Commission (1968) Chicago was hosting the most notorious
slums in the country.
4 Data
The Cook County Medical Examiner’s Officer13 has been reporting individual COVID-19
related deaths daily since March 16, 2020, the date of the first fatality recorded in the
county.14 The declared goal of the initiative is to provide direct access to critical facts
that can allow to identify communities that are most severely impacted by the virus. We
employ data recorded up to June 16, that we downloaded on June 18.
The Medical Examiner’s Office reports those deaths that are under its jurisdiction,
including among others those due to diseases constituting a threat to public health.15 The
information on COVID-19 fatalities coincides with that provided by the Johns Hopkins
University & Medicine Coronavirus Resource Center.16
The race, gender, age, comorbidities, and residence (home address, city, zip code,
and geographical coordinates) of each dead individual are also provided by the Medical
Examiner.17 Overall, 4,325 individuals—1,491 of whom black (that is, 34.47 percent)—
13See Cook County Medical Examiner COVID-19 Related Deaths athttps://datacatalog.cookcountyil.gov/Public-Safety/Medical-Examiner-Case-Archive-COVID-19-Related-Dea/3trz-enys.
14Data refer to deaths for which COVID-19 is reported among either primary or secondary causes.Operationally, the Medical Examiner’s Officer looks for references to COVID-19 in any of these fields:Primary Cause, Primary Cause Line A, Primary Cause Line B, Primary Cause Line C, or SecondaryCause.
15Namely, the Medical Examiner’s Office investigates any human death that falls within any or allof the following categories: criminal violence; suicide; accident; suddenly when in apparent good condi-tion; unattended by a practicing licensed physician; suspicious or unusual circumstances; unlawful fetaldeath as provided in Public Act 101-0013 of the 101st General Assembly of Illinois; poisoning or at-tributable to an adverse reaction to drugs and/or alcohol; disease constituting a threat to public health;injury or toxic agent resulting from employment; during some medical diagnostic or therapeutic proce-dures; in any prison or penal institution; when involuntarily confined in jail prison hospitals or otherinstitutions or in police custody; when any human body is to be cremated, dissected or buried at sea;when a dead body is brought into a new medico-legal jurisdiction without proper medical certification.Overall, each year, more than 16,000 deaths are reported to the Cook County Medical Examiner. Seehttps://www.cookcountyil.gov/agency/medical-examiner.
16Data may temporarily differ from those provided by the departments of public health because oftime lags in notification.
17Data are obtained from vital records, hospitals, and families.
139C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
020
4060
16mar2020 26mar2020 05apr2020 15apr2020 25apr2020 05may2020 15may2020 25may2020 04jun2020 14jun2020Week
Others Blacks
Figure 2: COVID-19 Deaths, by Race - Cook County, March 16-June 16, 2020Note: The figure reports the number of COVID-19 related deaths by day, separately for blacks and all other races combined.
have died of COVID-19 in Cook County between March 16, 2020 and June 16, 2020.
However, geographical coordinates are missing for 704 individuals. This leaves us with a
sample of 3,621 deaths—1280 (35.35 percent) of blacks—recorded over the over 14-week
span.18
Figure 2 plots the number of COVID-19 deaths in our sample for each day, separately
for blacks and all other races combined. From March 16 to April 9 the daily number
of blacks dying from COVID-19 is above the number for all other races, although the
share of the black population is only about one fourth of the total population. By April
9, the cumulative share of blacks who have died from COVID-19 represents almost 58
percent of the total COVID-19 deaths. The daily number of deaths among blacks keeps
increasing until mid-April and then starts decreasing at a slow pace. The number of
deaths among other races, on the other hand, keeps increasing until mid-May. By May
16, the cumulative share of black COVID-19 deaths is down to about 39 percent, to
decrease further to 35.3 percent by June 16. In other words, cumulatively to June 16,
18The racial distribution remains very similar, whether or not unreferenced individuals are included.
140C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
blacks in our sample are dying at a rate 1.3 times higher than their 27 percent sample
share in the population.
The above data clearly show that blacks are overrepresented in terms of COVID-
19 deaths. Furthermore, the data document that blacks likely became infected, and
eventually died, before the rest of the population, with a consequent decline in the share
of cumulative black deaths as the epidemic followed its course—a trend that has been
overlooked by media and public bodies reports.19
Figure A2 shows the distribution of deaths for blacks and all other races combined,
by age group.20 Blacks display a much larger number of deaths in their 60s and 70s,
and a lower number past their 80s. Nevertheless, the age distribution below age 60 is
quite similar. Overall, the graph confirms a large number of fatalities within the elderly
population, a phenomenon which has been largely documented. Figure A3 shows the
breakdown by gender among the two groups. Compared to other groups, blacks have a
much higher probability of death among women (almost 48 percent against 39 percent).21
We extract information on comorbidities by generating a set of 14 dummy variables
that take value one (and zero otherwise) when an individual who died from COVID-19 was
affected by diabetes and/or asthma, liver disease, cancer, hypertension, kidney disease,
obesity, respiratory diseases, neuro-cardiac diseases, neuro-respiratory diseases, asplenia,
immunodeficiency, transplant, and heart diseases.22 Figure A4 shows the distribution
of deaths by comorbidity and race. Diabetes and hypertension are by far the two most
common comorbidities. For both, blacks are more likely to suffer from them than the
other races combined.
We spatially merge the death data from the Medical Examiner with census block group
boundary files23 and with the redlining maps produced by HOLC and georeferenced by
the University of Richmond.24 Figure 3 shows the result of the spatial merge. Each indi-
vidual COVID-19 death is mapped into a specific block group and HOLC area using the
georeferenced home address of the deceased. The map highlights block group boundaries,
19On April 7, on the basis of the Medical Examiner’s data, the Chicago Tribune (Reys et al. 2020)—echoed by the Journal of the American Medical Association (Yancy, 2020) and the Lancet (Bhala,2020)—reported that 68 percent of the dead in the City of Chicago involved African Americans, whorepresent about 30 percent of the city’s population. As of June 16, African Americans account for lessthan 42 percent of the deaths in Chicago. Thus, the same trend can be detected both for the City ofChicago and Cook County as a whole.
20Only four deaths are reported below age 20, for three blacks aged 19, 18 and below 1, respectively,and a white aged 12. Figures are not normalized by the size of the population in each age group.
21Again, figures are not normalized by the degree of feminization of the population.22Disease groupings followed those employed by OpenSAFELY Collaborative (2020). Groups are not
mutually exclusive.23A block groups represents a combination of census blocks and a subdivision of a census tract. A
block group is defined to contain between 600 and 3,000 individuals.24See Nelson et al. (2020) and https://dsl.richmond.edu/panorama/redlining/.
141C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 3: Cook County Map and Total COVID-19 Deaths, March 16-June 16, 2020Note: The map reports census block group boundaries and HOLC areas, with green, blue, yellow, and red denoting re-spectively grade A, B, C, and D neighborhoods.
while HOLC areas are identified by the colour, with green, blue, yellow, and red denoting
respectively grade A, B, C, and D.
From the Cook County Government25 we also obtain census tract level data on socioe-
conomic characteristics (i.e., age, education, personal income, unemployment, population,
and racial groups), averaged over the period 2014-2018. We match these data using for
each block group the information available for the corresponding tract.
Figure A5 illustrates the share of black deaths by HOLC grade, only for census tracts
where a COVID-19 death has been reported. In D-ranked neighborhoods blacks represent
the highest share of the dead, while no black death is reported in those ranked A. Figure
A6 illustrates the share of blacks in the population by HOLC grade. With the only
exception of A-ranked neighborhoods, the share of the black population is lower than
the corresponding share of black deaths shown in Figure A5. This preliminary evidence
supports a role for redlining in determining segregation.
We complement the dataset with distance (measured in degrees) from a hospital and
25See Cook County Government Open Data at https://datacatalog.cookcountyil.gov/GIS-Maps/2010-U-S-Census-Mail-Return-Rates-and-Demographics/mpyu-4jqk. Data come from the American Commu-nity Survey estimates and are averages over the period 2014-2018 available by census tract
142C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
a nursing home.26 We interpret the latter as a proxy for the likelihood that a COVID-19
death occurred in a nursing home.27
Lastly, for the heterogeneity analysis in Sub-section 8.2, we employ the CDC’s So-
cial Vulnerability Index (SVI) dataset,28 created by the Geospatial Research, Analysis &
Services Program (GRASP) run by the Agency for Toxic Substances and Disease Reg-
istry (ATSDR) with the scope of helping emergency response planners and public health
officials in the face of a hazardous event, such as a natural disaster (e.g., a tornado or
epidemic) or a human-made event (e.g., an oil spill). Social vulnerability refers to those
factors that may affect the resilience of a geographical area to such an event. The source
of the data used to obtain these indices is the American Community Survey, that pro-
vides averages for geographic areas over the period 2014-2018 at a census tract level. The
SVI dataset include 15 characteristics, grouped into four indices: Socioeconomic Status
(comprising income, poverty, employment, and education variables), Household Com-
position/Disability (comprising age, single parenting, and disability variables), Minority
Status/Language (comprising race, ethnicity, and English language proficiency variables),
and Housing/Transportation (comprising housing structure, crowding, and vehicle access
variables). A general measure of social vulnerability, which summarizes the above four
indices, is also provided.29
Tables A1 and A2 in the Appendix reports variable definitions and sources, for the
cross section and the panel respectively.
5 Cross-sectional evidence
We start our analysis by exploiting cross-sectional information about individual deaths
from COVID-19 occurring in Cook County between March 16, 2020 and June 16, 2020.
Summary statistics for the cross-sectional sample are reported in Table A3.
A preliminary warning is in order, since the sample is clearly self-selected, as it in-
cludes only individuals who have died of COVID-19 and therefore exhibiting specific
characteristics, which are precisely those that tend to be more prevalent among blacks.
26We take georeferenced hospital location from the Cook County Government (see Cook County Healthand Hospitals Facilities at https://datacatalog.cookcountyil.gov/Economic-Development/Cook-County-Health-and-Hospitals-Facilities/jdix-z6uf) and georeferenced nursing home location from the MedicareNursing Home Compare dataset (see https://data.medicare.gov/data/nursing-home-compare).
27The Illinois Department of Public Health estimates that about 40 percent of the COVID-19 deathsin Cook County occurred in a long-term facility. See http://www.dph.illinois.gov/covid19/long-term-care-facility-outbreaks-covid-19.
28See https://svi.cdc.gov/data-and-tools-download.html and Flanagan et al. (2011).29Some of the variables, e.g., personal income and unemployment, coincide with those obtained through
the Cook County Government from the same source, i.e., the American Community Survey.
143C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Keeping in mind the sample selection problem afflicting the cross-sectional sample, in
order to assess whether residence in a historically redlined area is a predictor of the
probability that an individual that dies from COVID-19 is black, we employ as the
outcome variable a dummy taking value one if the individual that died from COVID-
19 in Cook County between March 16 and June 16 is reported to be black, and zero
otherwise. Table A3 shows that the probability of a black death is 35.3 percent.
Table A3 also indicates, for each HOLC area, the probability that a COVID-19 death
has occurred in that area. For instance, the probability that a COVID-19 death has
occurred in A-graded neighborhoods is 0.2 percent.30 Mean age is between 60 and 70
and 42 percent of the dead are females. Among pre-existing conditions, the disease with
highest prevalence is hypertension, that affects 52 percent of the sample, followed by
diabetes with nearly 41 percent.
The empirical model aims at exploiting the cross-sectional variation in the mortality
of blacks across HOLC areas. Formally, we estimate the following model:
Di,d = λd +5∑
h=1
βi1 (Hi = h) +X ′iπ + Z ′cρ+ µi,d (1)
where Di,d is a dummy taking value one if individual i that died from COVID-19 in day d
is black (and zero otherwise); λd represent day fixed effects that are meant to capture the
daily variation in the number of deaths;5∑
h=1
1 (Hi = h) are a full set of dummies denoting
the four HOLC-graded areas, from A to D, plus the ungraded area, where individual
i used to reside; the vector X ′i includes a set of individual characteristics (age group,
gender, comorbidities, and distance from the closest hospital and nursing home); the
vector Z ′c includes a set of socioeconomic characteristics at tract level (the logarithmic of
population and the shares of the population of age 18-64, over 65, without a high school
diploma, and black); µi,d is the error term which we cluster at day level.
Table 1 reports OLS estimates for eight variants of Equation 1. In Model 1 we only
control for HOLC (the omitted area is the ungraded area, No HOLC) and day fixed
effects. In Model 2 we add demographic information on age and gender, in Model 3
comorbidities, in Model 4 population, in Model 5 the shares of the population aged 18-64
and over 65, in Model 6 the share of the population without a high school diploma, and
in Model 7 distance from hospital and nursing home. In all models, the fact that an
individual that died from COVID-19 used to live in HOLC areas D and C is positively
and significantly associated with the probability that the individual is black, while the
30The reason why probabilities across the four HOLC areas do not sum up to one is that someneighborhoods of Cook County were not mapped by HOLC. 40 percent of the COVID-19 deaths occurredin ungraded neighborhoods of the county.
144C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 1: Black COVID-19 Death, Cross-Sectional Results - Cook County, March 16-June16, 2020
Black COVID-19 Death(1) (2) (3) (4) (5) (6) (7) (8)
HOLC A -0.3808*** -0.3652*** -0.3244*** -0.3261*** -0.3247*** -0.3636*** -0.3514*** -0.1250***(0.0640) (0.0720) (0.0757) (0.0800) (0.0811) (0.0798) (0.0751) (0.0424)
HOLC B 0.0389 0.0360 0.0399 0.0373 0.0355 0.0303 0.0072 0.0169(0.0329) (0.0334) (0.0327) (0.0323) (0.0327) (0.0322) (0.0316) (0.0250)
HOLC C 0.0800*** 0.0817*** 0.0830*** 0.0677*** 0.0716*** 0.1176*** 0.0848*** -0.0146(0.0177) (0.0185) (0.0181) (0.0175) (0.0202) (0.0216) (0.0234) (0.0182)
HOLC D 0.1752*** 0.1769*** 0.1745*** 0.0833*** 0.0852*** 0.1232*** 0.0922*** 0.0106(0.0233) (0.0238) (0.0231) (0.0255) (0.0266) (0.0281) (0.0297) (0.0222)
Age Groups No Yes Yes Yes Yes Yes Yes YesFemale No Yes Yes Yes Yes No Yes YesComorbidities No No Yes Yes Yes Yes Yes YesPopulation (log) No No No Yes Yes Yes Yes YesShare Aged 18-64 No No No No Yes Yes Yes YesShare Aged 65+ No No No No Yes Yes Yes YesShare With No High School No No No No No Yes Yes YesDistance From Hospital No No No No No No Yes YesDistance From Nursing Home No No No No No No Yes YesBlack Share No No No No No No No YesDay FE Yes Yes Yes Yes Yes Yes Yes Yes
Adj.R-squared 0.056 0.069 0.090 0.111 0.111 0.132 0.144 0.475Observations 3618 3616 3616 3616 3616 3616 3616 3616
Note: The dependent variable is a dummy variable that takes value one if an individual who died from COVID-19 is black,an zero otherwise. The omitted area is the one ungraded by the HOLC. Robust standard errors clustered at a day level inparentheses: *** p<0.01, ** p<0.05, * p<0.1.
opposite is true for area A. In other words, black mortality is much larger in low-graded
areas. To be noticed is that the effect is statistically equally significant for areas C and D,
even though the size of the coefficient is larger for D.31 In terms of magnitudes, relative to
the sample mean (35.3), the probability that an individual that died from COVID-19 is
black is over 24 percent larger in area C and 26 percent larger in area D. However, when
in Model 8 we also add the share of blacks in the census tract, having been a resident in
such areas is no longer a predictor of the dead individual’s racial group, suggesting that
indeed historical redlining induced segregation along racial lines, that fully absorbs its
influence in the present day. In other words, the black share fully captures the effect of
residence in a low-graded area, because of the correlation between black share and the
latter.
It is instructive to report how the dependent variable is affected by other covariates (all
coefficients are reported in Table A4). For instance, even in the full specification where
we control for the black share, its likelihood is higher for females, which means that,
within the black population, women have a relatively higher chance to die from COVID-
19, relative to the chance they have within the total population. Furthermore, pre-
existing conditions, in particular hypertension and kidney and respiratory diseases among
those with high prevalence, exert a significantly positive effect on the dependent variable.
Distance from hospital decreases the probability of a black death although it becomes
31The similarity between the effects of red and yellowlining is consistent with the evidence presentedby Aaronson et al. (2017) over a sample of US cities.
145C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
insignificant once we control for the share of blacks. Distance from a nursing home, that is,
a lower probability that a death has occurred in a nursing home, increases the probability
of a black death but, once race is accounted for, the sign of the association is reversed,
which can be attributed to a negative correlation between the presence of nursing homes
and the black share of a neighborhood. Nevertheless, neither comorbidities nor the other
demographic and socioeconomic factors fully absorb the influence of redlining, at least
until the black share is accounted for.
Overall, the cross-sectional results confirm that yellow and redlined areas are associ-
ated with a higher incidence of COVID-19 deaths among blacks, and that the difference
between redlined and yellowlined areas is blurred. However, since blacks are concen-
trated in the areas where mortality is higher, the cross-sectional analysis solely based on
information about those that died from COVID-19 is severely biased because of sample
selection. This limitation motivates the event study approach we introduce in the next
section.
6 An event study approach
6.1 Empirical strategy
The cross-sectional approach only provides simple correlations with no causal implica-
tions because of issues related to sample selection and omitted variable bias, which may
confound pre-existing differences between areas with and without COVID-19 mortality.
To reduce such biases, we assemble a weekly balanced panel at block group level over
the period from January 1, 2020 to June 16, 2020.32 For these 24 weeks, we collect
information on all types of deaths, that is, from COVID-19 and other causes under the
jurisdiction of the Cook County Medical Examiner’s Officer. Between January 1 and
June 16, 6,753 deaths—of which 5,492 after March 1633—are reported, each associated
with home residence, geographical coordinates, and individual characteristics.34 Again
32The merged dataset includes over 3,992 block groups and 1,318 census tracts.33This implies that between March 16 and June 16 there are more than 1,871 deaths in addition to
COVID-19 related deaths, which is consistent with the 1,256 deaths reported before March 16.34As we examine how mortality evolves before and after the epidemic outbreak, we should keep in mind
that, generally speaking, officially reported COVID-19 deaths do not match excess deaths, as measuredby the gap between observed deaths after the epidemic outbreak and deaths observed during the sameperiod in normal years. However, since our source is the Medical Examiner, the discrepancy betweenrecorded COVID-19 and excess deaths is greatly alleviated, because of the specific nature of the deathsunder his jurisdiction. Consequently, the deaths data we use in the cross-sectional analysis and those weuse in the event study are directly comparable, provided that it remains possible that deaths from othercauses under the Medical Examiner’s jurisdiction may also have risen or declined for various reasons(e.g., more individuals may have died because they did not receive care for other diseases, but fewer mayhave died for car accidents due to the lockdown). Indeed the Medical Examiner reports, to June 16,
146C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Figure 4: Cook County Map and Total Deaths, January 1-June 16, 2020Note: The map reports census block group boundaries and HOLC areas, with green, blue, yellow, and red denoting re-spectively grade A, B, C, and D neighborhoods.
we map individual deaths into census block groups and HOLC areas and then we aggre-
gate them at a census block group and by week. Therefore, for each block group-week,
we gather information on reported number of deaths (if any) for a given block group in
any of the 24 weeks from January 1 to June 16. Figure 4 illustrates the spatial merge.
The data is then merged with the previously described socioeconomic characteristics pro-
vided at census tract level by the Cook County Government and with proxies for social
vulnerability to shocks from the CDC.
Aggregating at a block group level has clear advantages in terms of identification,
since it allows to test for pre-treatment differences while controlling for block group fixed
effects which should filter out the effect of all socioeconomic factors which do not change
over this 24-week period. However, the aggregation implies that a given block group
may overlap multiple HOLC neighborhoods, possibly assigned to different grades. Figure
A7 zooms into the map and displays as an example four block groups, each including
neighborhoods that belong to three different HOLC types. In each block group, two
neighborhoods are graded C and outlined in yellow, and one is ungraded (the white
2020, over 3,000 excess deaths compared with the same period in 2019, a figure that closely matches theCOVID-19 deaths that were recorded.
147C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
0.0
5.1
.15
.2M
orta
lity
Rat
e
2020w1 2020w5 2020w9 2020w14 2020w18 2020w22Week
Others Blacks
Figure 5: Mortality Rate, by Race - Cook County, January 1-June 16, 2020Note: The figure reports mortality rates from any cause of death by week, separately for blacks and all other races com-bined.
one). The distribution of HOLC neighborhoods by block group is shown in Figure A8.
Almost 80 percent of the block groups in our sample include no more than two HOLC
neighborhoods (potentially with the same HOLC grade) and almost 95 percent of the
block groups include no more than three, with the rest of the block groups including up
to ten different HOLC neighborhoods. Thus, in order to determine treated block groups,
we will focus on those for which the majority of the surface area falls into HOLC areas
graded either C or D. This is consistent with the cross-sectional evidence, according
to which redlined and yellowlined areas exert a similar effect on the probability that
an individual that died from COVID-19 is black, and also with the evidence reported by
Aaronson et al. (2017), who stress the relevance of yellowlining for racial segregation. The
treatment therefore will capture whether resilience to external shocks, namely, COVID-
19, is affected by the historical policies which have favored racial segregation and economic
discrimination.
Figure 5 plots the mortality rate from any cause, separately for blacks and all other
148C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
races combined, for each week in the sample.35 From the beginning of the sample period,
the mortality rate is higher for blacks. For the first 10 weeks, before the epidemic out-
break, average mortality for blacks fluctuates between 0.04 and 0.05 deaths per thousand
versus around 0.02 per thousand for other races. Starting from the eleventh week (i.e.,
the week of March 13, that is the week when the first COVID-19 death is recorded on
March 16), mortality soars among both groups, but much more steeply so for blacks.
Both curves seem to be reaching a plateau by week 18 and start converging toward to
pre-COVID mortality rates by week 24.
Summary statistics for the panel dataset, including socioeconomic covariates, are re-
ported in Table A5. On average, for each block group-week in the sample, 0.070 deaths—
0.029 (41.4 percent) of blacks—are recorded during the period under examination. The
probability of a death in a given block group-week (measured as a dummy variable) is 5.7
percent, while the probability of a black death is 2.5 percent. Treated block groups (i.e.,
those predominantly graded C or D) represent almost 62 percent of the sample. Social
vulnerability indices represent percentile ranking of geographical areas depending on 15
categories and are bounded between 0 and 1, with 1 representing the maximal level of
vulnerability (the last percentile). A socioeconomic vulnerability of 0.54 therefore de-
notes that the average census tract within Cook County falls approximately within the
50th percentile of the distribution of vulnerability.
Figure A9 summarizes a number of demographic and economic characteristics by
HOLC grade. Personal income in red and yellowlined neighborhoods is well below $40,000
and even lower in the latter. A similar pattern emerges for the share of population with
no high school diploma, which is highest in C-graded neighborhoods. The share of the
black population is highest in redlined neighborhoods (close to 40 percent), while the rate
of black mortality peaks in yellowlined ones, with an average of two deaths per thousand
of blacks. However, in reading mortality statistics one must consider the fact that the
denominator (the black population) is not constant across the four areas. Figure A10
plots the weekly mortality rate for blacks and the other races by HOLC grade. Mortality
increases with the outbreak of the COVID-19 epidemic in all areas. The spikes we ob-
serve for black mortality in A-graded neighborhoods are due to the very small share of
the black population. In neighborhoods belonging to the other three areas, black mor-
tality is always higher than the one for other races and particularly high in yellowlined
areas, although one must consider that the black population in yellowlined areas is much
smaller than in redlined ones, as shown in the previous figure.
Our goal is to capture the impact of the asymmetric shock introduced by COVID-19 on
35To compute mortality rates, i.e., number of deaths over population, we first sum weekly deaths at acensus tract level (because population data is at a tract level) and then we collapse by week.
149C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
historically segregated areas, that is, whether treated block groups react differently after
treatment initiation. In other words, we aim at establishing whether, after the outbreak
of the epidemic, in majority C and D block groups deaths deviate more significantly from
those recorded in the pre-treatment period. Thus, we estimate variants of the model
below:
Yi,t = δi + γt +13∑
k=−10
βk1 (Ki,t = k) + εi,t (2)
where Yi,t represents the number of deaths in block group i and week t;36 δi and γt are
block group and week fixed effects meant to control for block group characteristics which
are fixed over the 24-week period and to capture natural fluctuations in mortality as well
as policies that affect the county uniformly (namely, the lockdown); 1(Ki,t = k) denote
treated block groups for ki,t = −10,−9, ..., 13 periods (i.e., weeks) before and after the
treatment kicks in, with the βks for k < 0 corresponding to pre-treatment effects and
the βks for k ≥ 0 corresponding to the dynamic effects k periods relative to the event.
We omit the period before the treatment kicks in, i.e., week 10. The error term, εi,t, is
clustered at block group level.
The event study approach outlined in Equation 2 allows to alleviate some of the
shortcomings affecting the cross-sectional analysis. The inclusion of the pre-treatment
periods allows to test for the parallel trend assumption and therefore for potential secular
differences between treated and control groups, as well as the occurrence of self-selection,
that may lead to different rates of disease transmission between treated and untreated
group. Issues related to sample selection will also be ruled out, since the sample includes
the universe of the block groups in the county. Learning and adaptation to the treatment
before it kicks in, as it occurs with staggered treatment, is also unlikely given that the
treatment period is constant. The only potential source of bias is therefore likely to
be related to measurement error, since not all the deaths that occur in Cook County
are reported to the Medical Examiner. Measurement error may be more severe in areas
with higher mortality but, since this is likely to cause an attenuation bias, the resulting
estimator would produce more conservative estimates.
36The alternative would be to use the mortality rate as a dependent variable. However, we haveracially-disaggregated data on population only at a census tract level rather than at block group level.In addition, block group fixed effects should absorb differences in population, given that the latter shouldremain constant over the span of only 24 weeks. In other words, using mortality data would amount todivide deaths by a constant, which should be picked up by the fixed effect.
150C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.05
0.0
5.1
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
a. Blacks
-.05
0.0
5.1
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
b. Others
Figure 6: Dynamic Effect of the Treatment on Deaths, by Race - Cook County, January1-June 16, 2020Note: The dependent variable is number of deaths, of blacks (Panel a) and of other groups (Panel b). The coefficientsare least-squares estimates of the βks. Block group and week fixed effects are included. Vertical lines represent 95 percentconfidence intervals based on standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
6.2 Baseline results
As shown in Figure 1, Cook County was only partially mapped by the HOLC in the
1930s, since the Corporation focused on cities with a population in 1930 above 40,000.
It is therefore possible that those neighborhoods which were mapped had completely
different characteristics from those which were not, and that these differences in pre-
existing conditions may be somewhat correlated with the treatment. To minimize the
possibility of biases arising from comparing areas that were already different, we focus
on block groups within municipalities which were mapped by the HOLC.37
Figure 6 illustrates the dynamic treatment effect on the number of deaths for blacks
(Panel a) and all other races (Panel b). The coefficients are least-squares estimates of the
βks and vertical lines represent 95 percent confidence intervals based on standard errors
clustered at at block group level. Before the COVID-19 outbreak, the average number of
37A robustness check that includes ungraded neighborhoods actually leads to even sharper results,reported in Figure A11 that are omitted for brevity and available upon request.
151C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
deaths for blacks and all other races in treated block groups is not significantly different
from the average number of deaths in non-treated ones. However, after the epidemic
shock, mortality in treated block groups increases quite sharply and this is particularly
evident among African Americans, for whom the number of deaths in treated block groups
increases up to 0.05 per week, an over 17 percent increase with respect to the average
number of black deaths by week and by block group. The effect starts picking up in
period 4 (toward mid April, when overall mortality sharply increases). Four weeks after
the outbreak, the number of deaths among African Americans increases by almost 0.03
per week and continues rising for the next few weeks before it starts fading away toward
the end of the 24-week period. By contrast, deaths for other races are not significantly
affected by the treatment.38
To quantify the overall effect of the treatment, in Table A6 we estimate the Average
Treatment Effect on the Treated (ATET) for blacks and the other races using a simple-
difference-in differences method. Overall, we find that the number of black deaths in
treated block groups after the epidemic shock (relative to non-treated blocks) increases
by 0.015, while the effect for the other races is not statically significant (and negative). If
compared to the average number of deaths for the untreated sample in the post-treatment
period (i.e., 0.025), the implied magnitude of the estimated effect is an increase in the
number of black deaths by almost 60 percent.
7 Robustness
7.1 Allowing for differential trends
The evidence presented in the previous section reveals an asymmetric effect of the epi-
demic shock and suggests that the relative resilience to it depends on the level of historical
racial segregation induced by loan market discrimination as a result of the HOLC Res-
idential Security Maps. An alternative explanation for the positive effect we detect for
the treatment points to potential differential trends which may have placed treated and
untreated neighborhoods on different trajectories that then surfaced only when the shock
struck. To control for such differential trends, we use tract level socioeconomic and de-
38Similar results obtain, as shown in Figure A12, if we replace the dependent variable with a dummyvariable capturing the occurrence of a death. This is also addressing the issue raised in footnote 36, sincethe definition of the dependent variable as a dummy does not reflect the distinction between numberof deaths and mortality rate. Other robustness checks which we do not report for brevity includeslightly changing the definition of the treatment group (Majority C & D may include some A and Bneighborhoods, so we also redefine the treatment group by excluding As, thus making the criterion morestringent) and using distributed and error lag models to respectively account for spillovers and serialcorrelation in the error.
152C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
mographic controls, that we interact with week dummies in order to provide the time
variation we aim at exploiting. The result of this exercise is reported in Figure A13 for
black deaths. In Panel a we interact with week dummies the share of the population aged
65 and above. In Panel b we further add interactions with the shares of Chinese speakers
and Hispanics. In Panel c we keep adding differential trends depending on unemployment
and income. In the last three plots we insert interactions with the share with no high
school diploma (Panel d), the share of blacks (Panel e), and distance from hospital and
nursing home (Panel f). Differential trends in income and unemployment partially offset
the effect of the treatment, but we can still observe for it a sizeable (although diminished)
effect on the number of deaths.
7.2 Robustness to aggregation
As mentioned above, aggregating at a block group level allows us to tighten the identifi-
cation of the effects. However, the definition of the treatment may be blurred by the fact
that a block group can overlap multiple HOLC areas (as in the example shown in Figure
A7). To explore the extent of issues that may be raising due to our aggregation approach,
we perform a sensitivity test and split the sample between block groups including a single
HOLC neighborhood and block groups including up to two, up to three, and up to four
HOLC neighborhoods. Block groups that only include a single HOLC neighborhood are
not affected by treatment definition issues due to aggregation, since each of them falls
entirely within a given HOLC category. As a result, comparing results obtained for these
block groups with those for block groups that include up to two, three and four neighbor-
hoods will allow us to understand whether aggregation is truly a problem. Reassuringly,
the effect of the treatment does not change sensibly whenever we include blocks which
include multiple HOLC neighborhoods (Figure A14).
8 Heterogeneity
8.1 Heterogeneity by HOLC grade
In order to evaluate factors that are potentially correlated with the partitioning of cities by
the HOLC and may explain the effect we found, this section goes on to investigate how the
effect of the treatment varies with specific characteristics. We start with a heterogeneity
analysis by HOLC grade. Specifically, we consider sub-samples of neighborhoods, defined
on the basis of their HOLC ranking.
153C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.1-.0
50
.05
.1Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
a. HOLC D vs A
-.1-.0
50
.05
.1Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
b. HOLC D vs B-.1
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
c. HOLC C vs A
-.1-.0
50
.05
.1Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
d. HOLC C vs B
Figure 7: Heterogeneity by HOLC Grade - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups for which the majority of the surface area falls in either a grade A or a grade D area (Panel a), ineither a grade B or a grade D area (Panel b), in either a grade A or a grade C area (Panel c), and in either a grade B ora grade C area (Panel d). Block group and week fixed effects are included. Vertical lines represent 95 percent confidenceintervals based on standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
Results are shown in Figure 7. In Panel a we focus only on block groups for which
the majority of the surface area falls in either a grade A or a grade D area.39 In other
words, we compare the best and the worst neighborhoods (according to the HOLC grading
scheme). In Panel b we compare the second best neighborhoods (HOLC grade B) with
the worst (grade D). In Panel c we compare the best neighborhoods (grade A) with the
second worst (grade C) and in Panel d the second best (grade B) with the second worst
(grade C). The effect of the treatment is much larger when we compare the worst and
the second worst neighborhoods with the best (respectively Panels a and c), and becomes
significant from period 2 in Panel a and from period 3 in Panel c. However, even when we
compare the worst and the second worst neighborhoods with the second best (respectively
Panel b and Panel d), we still observe a quite significant effect of the treatment for several
weeks.
39This means that we drop block groups for which the majority of the surface area falls in grade Band C areas.
154C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Socioeconomic V. Below Median
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. Socioeconomic V. Above Median
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Socioeconomic V. Below Median & Black Share<25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Socioeconomic V. Below Median & Black Share>25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Socioeconomic V. Above Median & Black Share<25%-.2
-.10
.1.2
.3Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Socioeconomic V. Above Median & Black Share>25%
Figure 8: Heterogeneity by Socioeconomic Status Vulnerability - Cook County, January1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks oversamples of block groups with socioeconomic status vulnerability below median (Panel a), above median (Panel b), belowmedian and with black share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d),above median and with black share below 25 percent (Panel e), and above median and with black share above 25 percent(Panel f). Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals basedon standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
8.2 Heterogeneity by social vulnerability
The social vulnerability indices collected at a census tract level by the CDC allow us
to carry out a heterogeneity analysis depending on the level of vulnerability to shocks
of specific areas. Thus, in this section we group neighborhoods on the basis of the four
distinct dimensions of social vulnerability, as well as some of their components.
Figure 8 shows heterogeneity results by socioeconomic status vulnerability, an index
that comprises four components, reflecting percentile scores for personal income, poverty,
unemployment, and education. In Panels a and b we split the sample between block
groups with a value of the socioeconomic vulnerability index below and above the median.
In the next two panels (Panels c and d) we split block groups below median socioeconomic
vulnerability between those with a population black share respectively below and above
155C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
25 percent.40 In the two bottom panels (Panels e and f) we repeat the same exercise for
block groups with above median socioeconomic vulnerability. Unsurprisingly, the effect
of the treatment is much larger (and significant) in block groups with above median
socioeconomic vulnerability (Panel b). The estimated dynamic treatment effect in the
sample of block groups below median is nearly zero and not significant (at a 5 percent
level), despite the fact that standard errors are exceptionally small. When we split
the two sub-samples between those with a share of blacks below (Panels c and e) and
above (Panels d and f) 25 percent, some noteworthy differences emerge. The treatment
effect is much larger in neighborhoods with a black share above 25 percent, which is not
surprising given that the black share is this sub-sample is larger. However, among these
neighborhoods, the treatment effect changes quite significantly depending on the level of
socioeconomic vulnerability (Panel d vs Panel f). For areas below median vulnerability
(Panel d), the post-treatment effect of being red or yellowlined is not that different from
the pre-treatment, and in any case not significantly so at a 5 percent level. When, on the
other hand, we look at the sample above median vulnerability, that is, at neighborhoods
with worse performances, we observe quite a strong treatment effect, which resembles
the one we found over the whole sample. These dissimilarities point to a much stronger
impact of socioeconomic determinants, rather than genetic and biological one, of black
mortality. In other words, neighborhoods with the same black share exhibit different
effects of the treatment, and therefore different level of resilience to external shocks,
depending on the level of socioeconomic vulnerability.
To better understand which component of the socioeconomic vulnerability index drives
our findings, in Figures A15-A18 we replicate the same analysis focusing, one by one, on
its four components, that is personal income, poverty, unemployment, and education.
Overall, the effect of the treatment is much stronger in neighborhoods that underperform
in all four dimensions and at the same time exhibit a black share above 25 percent.
However, when we focus on neighborhoods with a black share above 25 percent, the
difference between the worst and the best performing neighborhoods is especially striking
for the income and poverty dimensions. The treatment effect for neighborhoods below
median income (Figure A15, Panel d) is large and significant, while above median income
(Panel f) the treatment effect is not significant in most post-treatment periods. The same
occurs for poverty (Figure A16), with a relatively small (and significant in one period
only) average treatment effect in the post-treatment period for neighborhoods with a
population share below the poverty line smaller than the median (Panel d), and a sizeable
effect for those with a larger one (Panel f).
40We choose this threshold because it is close to the average black share in the county. Furthermore,a higher threshold would greatly reduce the size of the sub-sample displaying a relatively higher blackshare.
156C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. HH Composition V. Below Median
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. HH Composition V. Above Median
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. HH Composition V. Below Median & Black Share<25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. HH Composition V. Below Median & Black Share>25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. HH Composition V. Above Median & Black Share<25%-.2
-.10
.1.2
.3Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. HH Composition V. Above Median & Black Share>25%
Figure 9: Heterogeneity by Household Composition Vulnerability - Cook County, January1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with household composition vulnerability below median (Panel a), above median (Panel b), belowmedian and with black share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d),above median and with black share below 25 percent (Panel e), and above median and with black share above 25 percent(Panel f). Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals basedon standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
In Figure 9 we repeat the same analysis with a focus on the household composition
vulnerability index, which includes four components, that is the shares of the population
over 65, below 17, older than 5 with disabilities, and of single parents. Again the effect
is much more marked when we focus on block groups with above median household
vulnerability and a black share above 25 percent (Panels d and f). Once again, to
understand which component of the index drives the effect, in Figure A19 we split the
sample between block groups with a share of population aged 65+ below and above
the median, while in Figure A19 we do so according to the share of single parents. As
expected, aging is quite an important factor. Consistent with the pattern reported for
blacks in Figure A2, fatalities among the elderly are more numerous when we look at
the sample with a black share above 25 percent (Panels d and f). For the share of single
parents in Figure A20, the difference in the estimated effect between Panel d and Panel
157C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Minority V. Below Median
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. Minority V. Above Median
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Minority V. Below Median & Black Share<25%
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Minority V. Below Median & Black Share>25%
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Minority V. Above Median & Black Share<25%-.4
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Minority V. Above Median & Black Share>25%
Figure 10: Heterogeneity by Minority Status Vulnerability - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with minority status vulnerability below median (Panel a), above median (Panel b), below median andwith black share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d), above me-dian and with black share below 25 percent (Panel e), and above median and with black share above 25 percent (Panel f).Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals based on standarderrors clustered at at block group level. The omitted period k = −1, i.e., week 10.
f is sizeable. In other words, the effect of yellow and redlining is most detrimental when
a high share of single parents is combined with a high black share.
In Figure 10 we repeat the heterogeneity analysis by vulnerability depending on mi-
nority status (the index also comprises a measure of English language proficiency) and
in Figure 11 by housing/transportation vulnerability (the index comprises percentages
of multi-unit structures, mobile homes, a measure of crowding based on the presence of
more people than rooms, households with no vehicle, and households in group quarters).
Differences in the treatment depending on vulnerability along these two dimensions, and
combined with the black share, are still detectable but less pronounced than those we
found for socioeconomic status and household composition vulnerability.
158C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Housing V. Below Median
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. Housing V. Above Median
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Housing V. Below Median & Black Share<25%
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Housing V. Below Median & Black Share>25%
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Housing V. Above Median & Black Share<25%-.4
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Housing V. Above Median & Black Share>25%
Figure 11: Heterogeneity by Housing/Transportation Vulnerability - Cook County, Jan-uary 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with housing/transportation vulnerability below median (Panel a), above median (Panel b), belowmedian and with black share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d),above median and with black share below 25 percent (Panel e), and above median and with black share above 25 percent(Panel f). Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals basedon standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
8.3 Heterogeneity by nursing home location
Lastly, the high levels of mortality recorded in nursing homes have been stressed both by
the media and the medical literature. To explore the relevance of this potential channel,
using the information on nursing home location available from Medicare41 we generate
the minimal centroid distance of a block group from a nursing home and we replicate the
heterogeneity analysis focusing on block groups located within 0.009 degrees (i.e., 1 km)
from a nursing home. Our aim is to capture the probability that a recorded death has
occurred in a nursing home. Figure 12 shows that the treatment effect is much stronger
in neighborhoods further away from nursing homes and that this is especially true when
the black share is above 25 percent. Thus, we can confidently exclude potential concerns
related to the possibility that we could have captured the effect of deaths in nursing
41See https://data.medicare.gov/data/nursing-home-compare.
159C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Distance Nursing Home<1km
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. Distance Nursing Home>1km
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Distance Nursing Home<1km & Black Share<25%
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Distance Nursing Home<1km & Black Share>25%
-.4-.2
0.2
.4Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Distance Nursing Home>1km & Black Share<25%-.4
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Distance Nursing Home>1km & Black Share>25%
Figure 12: Heterogeneity by Distance from Nursing Home - Cook County, January 1-June16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with distance from nursing home below 1 km (Panel a), above 1 km (Panel b), below 1 km and withblack share below 25 percent (Panel c), below 1 km and with black share above 25 percent (Panel d), above median andwith black share below 25 percent (Panel e), and above median and with black share above 25 percent (Panel f). Blockgroup and week fixed effects are included. Vertical lines represent 95 percent confidence intervals based on standard errorsclustered at at block group level. The omitted period k = −1, i.e., week 10.
homes.
9 The Hispanic population
While our main focus so far has been on how the African American population has been
hit by COVID-19, the Hispanic population has also been the subject of concern, both in
the media and the medical literature.42 Therefore, in this section, we extend the previous
analysis of COVID-19 outcomes to the white Hispanic population of Cook County.43
42See, for instance, Singh and Koran (2020) and Yancy (2020).43The Medical Examiner’s racial classifications are based on US Census Bureau categories, according
to which Latino, or equivalently Hispanic, is defined as ethnicity, and can therefore belong to any racialgroup. This section focuses on white Hispanics, who in Cook County represent the vast majority ofHispanics. To June 16, only 13 deaths were reported for black Hispanics, i.e., 0.09 percent of the black
160C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
The history of Latino immigration to Cook County starts at least as early as in the pe-
riod 1916-1928, when a steady and large flow of Mexicans moved to Chicago to find work
in the railroad and steel industries. Another wave took place in the period 1942-1964.
Their settlement pattern was similar to that of blacks and they were also historically
affected by segregation and redlining (Betancur, 1996). In fact Hoyt (1933) placed Mexi-
cans last, after blacks, in his ranking of the influence of ethnic groups on property values.
Figure A21 is a replica of Figure 2 that also reports Hispanic COVID-19 deaths,
as well as blacks and the remaining groups combined, from March 16 to June 16. Even
though the number of Hispanic deaths does increase in the initial weeks, it stays below the
number of black deaths. Figure A22 is a replica of Figure 5 that plots the Hispanic overall
mortality rate from January 1 to June 16. Strikingly, before the epidemic, the mortality
rate for Hispanic is smaller than that of blacks and even of other groups combined. It
does increase with the COVID-19 outbreak, but remains relatively contained if compared
to other groups.
The fact that Hispanics exhibit lower levels of mortality than the rest of the popula-
tion is actually a well-known fact, that has been referred to as the “Latino paradox”, as
Hispanics tend to display relatively favorable health outcomes despite their low socioeco-
nomic status (Markides and Coreil, 1986; Abraido-Lanza et al., 1999).
Nevertheless, when in Figure A23 we replicate Figure 6 with the baseline event study,
including results for Hispanics, we find that they are also affected by the treatment, even
though the effect kicks in later, and lasts longer, if compared to blacks. This confirms
that HOLC practices carry long-term implications, in terms of the induced resilience to
the epidemic shock, also for Hispanics.
10 Falsification tests
10.1 From 2017 to 2019
There are still additional potential threats to identification. To test the extent of such
threats, in this section we will carry out two sorts of falsification tests.
First, it might be the case that the treatment captures endemic annual trends in
mortality that would have anyway occurred and that have nothing to do with the COVID-
19 epidemic. To test the extent of such threat to identification, we carry out falsification
tests using the number of deaths in the corresponding time frame for the three years
prior to the epidemic, that is, from January 1 and June 16 in 2017, 2018, and 2019. The
deaths, and 1.6 percent of the Hispanic deaths.
161C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.04
-.02
0.0
2.0
4D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Blacks
-.04
-.02
0.0
2.0
4D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. Others
2019-.0
4-.0
20
.02
.04
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Blacks
-.04
-.02
0.0
2.0
4D
eath
s-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13
Weeks
b. Others2018
-.04
-.02
0.0
2.0
4D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Blacks
-.04
-.02
0.0
2.0
4D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. Others2017
Figure 13: Dynamic Effect of the Treatment on Deaths, by Race - Cook County, January1-June 16, 2019, 2018, and 2017.Note: The dependent variable is number of deaths, of blacks (Panels a) and of other groups (Panels b), in the periodfrom January 1 to June 16, in 2019 (top panels), 2018 (middle panels), and 2017 (bottom panels). The coefficients areleast-squares estimates of the βks. Block group and week fixed effects are included. Vertical lines represent 95 percentconfidence intervals based on standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
goal is to gauge the possibility that the treatment is capturing annual trends in mortality
related to the diffusion of other diseases (e.g., the flu).44 By replicating the analysis over
the previous years, we shall be able to ascertain that we are not merely capturing a yearly
trend having nothing to do with the COVID-19 epidemic.
Figure 13 shows that there is no effect of the treatment on black or other deaths in
either of the three previous years over the same time frame, which confirms that the effect
we found is attributable to the 2020 COVID-19 epidemic.
10.2 The Spanish flu
A further threat to identification is due to the fact the transmission rate of diseases (not
mortality) may have always been larger in redlined neighborhoods, so that the 1930s
44Data for the years 2017-2019 are also provided by the Medical Examiner’s Office.
162C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table 2: Spanish Flu - Chicago, 1918
Deaths From Spanish Flu(1) (2) (3) (4)
Majority C & D 0.3020* 0.1039 0.1257 0.1070(0.1588) (0.1525) (0.1498) (0.1479)
Population Density (log) Yes Yes Yes YesIlliteracy Rate No Yes Yes YesHome Ownership Rate No No Yes YesUnemployment Rate No No No YesWeek FE Yes Yes Yes Yes
Adj.R-squared 0.252 0.281 0.284 0.287Observations 3416 3416 3416 3416
Note: The dependent variable is the number of deaths from Spanish flu in Chicago in 1918. Robust standard errors clus-tered at a census tract level in parentheses: *** p<0.01, ** p<0.05, * p<0.1.
rankings may have merely formalized existing conditions. In other words, it may be the
case that low-graded neighborhoods may have already been subject to higher transmission
of viral diseases, even prior to redlining. To make sure that our results are not affected
by this kind of threat, we focus on Spanish flu mortality.
By 1918, Chicago had already experienced the first wave of the Great Migration, with
the black population rising from about 34,000 to 92,000 between 1910 and 1920, and
a parallel increase of racial segregation (Tuttle, 1970). Chicago was badly hit by the
Spanish flu, but blacks were actually hit less harshly than whites, as in the rest of the
country, a fact that is still largely unexplained (Gamble, 2010; Okland and Mamelund,
2019). From our perspective, the fact that blacks were already segregated, by and large,
in the same areas of the city, while at the same time they were somewhat protected from
the disease, suggests that those areas were not per se more unhealthy.
In order to perform a falsification test, we use census tract level data on Spanish flu
mortality in Chicago in 1918 to test whether the transmission of epidemics in low-graded
areas was already higher before they were assessed by the HOLC. Data are provided by
the Infectious Disease Dynamics Group.45 The dataset contains census tract location and
week of epidemic of 8,031 influenza and pneumonia deaths as well as sociodemographic
data (including population density, the illiteracy rate, the home ownership rate, and the
unemployment rate) for 496 census tracts within the City of Chicago. However, race-
disaggregated information is not reported.46
Table 2 reports four variants of a model where the number of Spanish flu deaths in
1918 is the dependent variable. In Model 1, where we only control for majority C or
D neighborhoods and week fixed effect, we find a marginally significant effect of the
45See http://www.ufiddynamics.org/data. The sources of the data are the 1920 Census and the 1920annual report of City of Chicago Department of Health.
46See Table A7 for summary statistics.
163C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
control variable on deaths. However, when in the following models we add (the log of)
population density, illiteracy, home ownership, and unemployment, no residual influence
remains. Thus, reassuringly, we find no evidence that the HOLC areas that we found
to be associated with higher COVID-19 mortality were subject to higher infections rates
prior to the implementation of the redlining policies of the 1930s.
11 Conclusion
Not only the United States is registering the worldwide highest number of fatalities from
the COVID-19 pandemic but, within the country, the death toll on African Americans
has been disproportionately large. Up to now, however, lack of race-disaggregated data
has prevented a rigorous assessment of this phenomenon and of its determinants.
Using so far unexploited individual and georeferenced death data collected by the Cook
County Medical Examiner, we provide first evidence that race does affect COVID-19
outcomes. The data confirm that in Cook County blacks are overrepresented in terms of
COVID-19 related deaths since—cumulatively in the period that goes from the outbreak
of the epidemic on March 16, 2020 until June 16, 2020—they constitute 35 percent of
the dead, which implies that they have been dying at a rate 1.3 times higher than their
population share.
Furthermore, by combining the spatial distribution of mortality with the redlining
maps for the Chicago area, we obtain a block group level panel dataset of weekly deaths
from all causes, over the period January 1, 2020-June 16, 2020, to which we apply an
event study design, where the treated neighborhoods are those that were historically
either yellow or redlined and treatment initiation coincides with the outbreak of the
COVID-19 epidemic. We show that, while no pre-treatment differences are detected, after
the outbreak of the epidemic on March 16, 2020 historically low-graded neighborhoods
display a sharper increase in mortality, which is driven by blacks. Thus, we uncover
a persistence influence of the racial segregation induced by the discriminatory lending
practices introduced in the 1930s.
We also establish that this influence runs by way of an asymmetric effect of the epi-
demic shock, which is in turn channeled through a diminished resilience of the black
population to the shock represented by the COVID-19 outbreak. Far from being deter-
mined by genetic and biological factors, such vulnerability can be linked to socioeconomic
status and household composition, as the likely channels through which the legacy of the
past manifests itself.
To conclude, one of the stylized facts emerging from this paper, and that deserves
164C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
further attention, is that not only blacks are disproportionately hit by COVID-19, but also
that they started to succumb to it earlier than other groups. Several explanations may be
behind this phenomenon. On the one hand, it is possible that blacks become infected as
much as the rest of the population, but they experience a faster progression through the
stages of the disease, because of pre-existing medical conditions and/or access to health
care. It may also be the case that blacks were more exposed from the beginning of the
outbreak, because of their occupations and living conditions, so that once stay-at-home
orders where issued they benefitted from them more thoroughly. The age and gender
composition of each racial group also may also play a role.47 The fact remains that the
evolution of the epidemiological curve reveals for the US an extraordinary degree of racial
and ethnic segregation, with different groups displaying profoundly distinct patterns even
in the timing of their exposure to the epidemic.
REFERENCES
Aaronson D., Hartley D., Mazumder B. (2017), The effects of the 1930s HOLC “redlin-
ing”maps. Federal Reserve Bank of Chicago Working Paper No. 2017-12.
Abraido-Lanza A.F., Dohrenwend B.P; Ng-Mak D.S., Turner J.B. (1999). The Latino
mortality paradox: A test of the “salmon bias”and healthy migrant hypotheses. American
Journal of Public Health, 89, (10): 15431548.
Alfani G., Murphy T. (2017), Plague and lethal epidemics in the pre-industrial World.
Journal of Economic History, 77, 314–343.
Almond, D. (2006). Is the 1918 influenza pandemic over? Long-term effects of in utero
influenza exposure in the post-1940 U.S. population. Journal of Political Economy, 114,
672–712.
Aassve A., Alfani G., Gandolfi F., Le Moglie M. (2020), Epidemics and trust: The
case of the Spanish flu. IGIER Working Paper No. 661.
Almagro M., Orane-Hutchinson A. (2020), The determinants of the differential expo-
sure to COVID-19 in New York City and their evolution over time. Covid Economics,
13, 31–50.
Anders J. (2019), The long run effects of de jure discrimination in the credit market:
How redlining increased crime. Mimeo.
Anderson S. (2018), Legal origins and female HIV. American Economic Review, 108,
1407–1439.
Appel I., Nickerson J. (2016), Pockets of poverty: the long-term effects of redlining.
Mimeo.
47In a companion paper (Bertocchi and Dimico, 2020b) we examine COVID-19 outcomes along thegender and work dimensions.
165C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Bertocchi G. (2016), The legacies of slavery in and out of Africa. IZA Journal of
Migration, 5, 1–19.
Bertocchi G., Dimico A. (2020a), Bitter sugar: Slavery and the black family. CEPR
Discussion Paper No. 14837.
Bertocchi G., Dimico A. (2020b), COVID-19, gender, and work. Mimeo.
Bertocchi G., Dimico A. (2019), The long-term determinants of female HIV infection
in Africa: The slave trade, polygyny, and sexual behavior. Journal of Development
Economics, 140, 90–105.
Bertocchi G., Dimico A. (2014), Slavery, education, and inequality. European Eco-
nomic Review, 70, 197–209.
Bertocchi G., Dimico A. (2012), The racial gap in education and the legacy of slavery.
Journal of Comparative Economics, 40, 581–595.
Betancur J.J. (1996), The settlement experience of Latinos in Chicago: Segregation,
speculation, and the ecology model. Source: Social Forces, 74, 1299–1324.
Bhala N., Curry G., Martineau A.R., Agyemang C., Bhopal R. (2020). Sharpening
the global focus on ethnicity and race in the time of COVID-19. Lancet. Comment. May
8.
Borjas G.J. (2020), Demographics determinants of testing incidence and Covid-19
infections in New York City neighborhoods. Covid Economics, 3, 12-39.
Boustan L.P. (2011), Racial residential segregation in American cities. In Brooks N.,
Donaghy K., Knaap G.-J. (eds.), Oxford Handbook of Urban Economics and Planning.
New York: Oxford University Press, 318–339.
Cage J., Rueda V. (2020), Sex and the mission: The conflicting effects of early Chris-
tian investments on sub-Saharan Africa’s HIV epidemic. Journal of Demographic Eco-
nomics, forthcoming.
Desmet K., Wacziarg R. (2020), Understanding spatial variation in COVID-19 across
the United States. CEPR Discussion Paper No. 14842.
Douglas Commission [National Commission on Urban Problems] (1968), Building the
American City. Washington: GPO.
Eligon J., Burch A.D.S., Searcey D., Oppel R.A.Jr. (2020), Black Americans face
alarming rates of coronavirus infection in some states. The New York Times, April 7.
Couch K., Fairlie R.W., Xu H. (2020), The impacts of COVID-19 on minority unem-
ployment: First evidence from April 2020 CPS microdata. IZA Discussion Paper No.
13264.
Federal Housing Administration (1938), Underwriting Manual: Underwriting and Val-
uation Procedure Under Title II of the National Housing Act. Washington: Federal
Housing Administration.
166C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Fishback P., Rose J., Snowden K. (eds.) (2013), Well Worth Saving: How the New
Deal Safeguarded Home Ownership. Chicago: University of Chicago Press.
Flanagan B., Gregory E., Hallisey E., Heitgerd J., Lewis B. (2011), A social vulner-
ability index for disaster management. Journal of Homeland Security and Emergency
Management, 8, 1–22.
Gamble V.N. (2010), “There Wasnt a Lot of Comforts in Those Days:”African Amer-
icans, Public Health, and the 1918 Influenza Epidemic. Public Health Reports, Supple-
ment 3, 125 , 114–121.
Greer J.L. (2014), Historic home mortgage redlining in Chicago. Journal of the Illinois
State Historical Society, 107, 204–233.
Greer J.L. (2012), Race and mortgage redlining in the United States. Mimeo.
Guimbeau A., Menon N., Musacchio A. (2020), The Brazilian bombshell? The long-
term impact of the 1918 influenza pandemic the South American Way. NBER Working
Paper No. 26929.
Harriss C.L. (1951), History and Policies of the Home Owners’ Loan Corporation.
New York: NBER.
Helgertz J., Bengtsson T. (2019), The long-lasting influenza: The impact of fetal stress
during the 1918 influenza pandemic on socioeconomic attainment and health in Sweden,
1968-2012. Demography, 56, 1389–1425.
Hillier A. (2005), Residential Security Maps and neighborhood appraisals: The Home
Owners Loan Corporation and the case of Philadelphia. Social Science History, 29, 207–
233.
Hillier A. (2003), Redlining and the Home Owners Loan Corporation. Journal of
Urban History, 29, 394–420.
Hoyt H. (1933), One Hundred Years of Land Values in Chicago: The Relationship of
the Growth of Chicago to the Rise of Its Land Values, 1830-1933. Chicago: University
of Chicago Press.
Hunt D.B. (2009), Blueprint for Disaster: The Unraveling of Chicago Public Housing.
Chicago: University of Chicago Press.
Institute of Medicine (2003), Unequal Treatment: Confronting Racial and Ethnic Dis-
parities in Health Care. Washington: National Academies Press.
Kendi, I.X. (2020), Stop looking away from the race of COVID-19 victims. The
Atlantic, April 1.
Jackson K. (1985), Crabgrass Frontier: The Suburbanization of the United States.
New York: Oxford University Press.
Jackson K. (1980), Race, ethnicity, and real estate appraisal: The Home Owners Loan
Corporation and the Federal Housing Administration. Journal of Urban History, 6, 419–
167C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
452.
Jedwab R., Johnson N.D., Koyama M. (2019), Negative shocks and mass persecutions:
Evidence from the black death. Journal of Economic Growth, 24, 345–395.
Jedwab R., Johnson N.D., Koyama M. (2016), Bones, bacteria and break points: The
heterogeneous spatial effects of the Black Death and long-run growth. GMU Working
Paper in Economics No. 16-30.
Karlsson, M., Nilsson T., Pichler S. (2014), The impact of the 1918 Spanish flu epi-
demic on economic performance in Sweden: An investigation into the consequences of an
extraordinary mortality shock. Journal of Health Economics, 36, 1–19.
Kirby T. (2020), Evidence mounts on the disproportionate effect of COVID-19 on
ethnic minorities. Lancet Respiratory Medicine. News. May 8.
Krieger N., Wright E., Chen J.T., Waterman P.D., Huntley E.R., Arcaya M. (2020),
Cancer stage at diagnosis, historical redlining, and current neighborhood characteris-
tics: Breast, cervical, lung, and colorectal cancer, Massachusetts, 2001-2015. American
Journal of Epidemiology, in press.
Krieger N., Van Wye G., Huynh M., Waterman P.D., Maduro G., Li W., Gwynn
R.C., Barbot O., Bassett M.T. (2020), Structural racism, historical redlining, and risk of
preterm birth in New York City, 20132017. American Journal of Public Health, in press.
Krimmel J. (2018), Persistence of prejudice: Estimating the long term effects of redlin-
ing. Mimeo.
Landrine H., Corral I., Lee J.G.L., Efird J.T., Hall M.B., Bess J.J. (2017), Residential
segregation and racial cancer disparities: A systematic review. Journal of Racial and
Ethnic Health Disparities, 4, 1195–1205.
Lang K., Kahn-Lang Spitzer A. (2020), Race discrimination: An economic perspective.
Journal of Economic Perspectives, 34, 68–89.
Lin, M.-J., Liu e.M. (2014), Does in utero exposure to illness matter? The 1918
influenza epidemic in Taiwan as a natural experiment. Journal of Health Economics, 37,
152–163.
Loper J. (2019), Womens position in ancestral societies and female HIV: The long-term
effect of matrilineality in sub-Saharan Africa. Mimeo.
McLaren J. (2020), Racial disparity in COVID-19 deaths: Seeking economic roots
with census data. NBER Working Paper No. 27407.
Markides K.S:, Coreil J. (1986), The health of Hispanics in the southwestern United
States: An epidemiologic paradox. Public Health Reports, 101, 253–265-
Nardone A., Casey J.A., Morello-Frosch R., Mujahid M., Balmes J.R., Thakur N.
(2020), Associations between historical residential redlining and current age-adjusted
rates of emergency department visits due to asthma across eight cities in California: An
168C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
ecological study. Lancet Planet Health, 4, e24–31.
Nelson R.K., Winling L., Marciano R., Connolly N., et al. (2020), Mapping inequality:
Redlining in New Deal America. In Nelson R.K., Ayers E.L. (eds.), American Panorama:
An Atlas of United States History. University of Richmond: Digital Scholarship Lab.
Nier, C.L.III (1999), Perpetuation of segregation: Toward a new historical and legal
interpretation of redlining under the Fair Housing Act. John Marshall Law Review, 32,
617–665.
Nunn N. (2008), Slavery, inequality, and economic development in the Americas: An
examination of the Engerman-Sokoloff hypothesis. In Helpman E. (ed.), Institutions and
Economic Performance. Cambridge: Harvard University Press, 48–180.
Okland H., Mamelund S.-E. (2019), Race and 1918 influenza Pandemic in the United
States: A review of the literature. International Journal of Environmental Research and
Public Health, 16, 1–18.
OpenSAFELY Collaborative (2020), OpenSAFELY: Factors associated with COVID-
19-related hospital death in the linked electronic health records of 17 million adult NHS
patients. medRxiv, May 7.
Oppel R.J.Jr., Gebeloff R., Lai K.K.R., Wright W., Smith M. (2020), The fullest look
yet at the racial inequity of coronavirus. The New York Times, July 5.
Orsi J.M., Margellos-Anast H., Whitman S. (2010), Black-white health disparities in
the United States and Chicago: A 15-year progress analysis. American Journal of Public
Health, 100, 349–356.
Platt L., Warwick R. (2020), Are Some Ethnic Groups more Vulnerable to COVID-19
than Others? London: Institute for Fiscal Studies.
Price-Haywood E.G., Burton J., Fort D., Seoane L. (2020), Hospitalization and mor-
tality among black patients and white patients with Covid-19. The New England Journal
of Medicine, 382, 2534–2543.
Reyes C., Husain N., Gutowski C., St. Clair S., Pratt G. (2020), Chicago’s coronavirus
disparity: Black Chicagoans are dying at nearly six times the rate of white residents, data
show. Chicago Tribune, April 6.
Richardson G., McBride M. (2009), Religion, longevity, and cooperation: The case of
the craft guild. Journal of Economic Behavior & Organization, 71, 172–186.
Rothstein R. (2017), The Color of Law: A Forgotten History of How Our Government
Segregated America. New York: Liveright.
Schill M.H., Wachter S.M. (1995), The spatial bias of federal housing law and policy:
Concentrated poverty in urban America. University of Pennsylvania Law Review, 143,
1285–1342.
Schmitt-Grohe S., Teoh K., Uribe M. (2020), COVID-19: Testing inequality in New
169C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
York City. Covid Economics, 8, 27-44
Shertzer A., Twinam T., Walsh R. (2016), Race, ethnicity, and discriminatory zoning.
American Economic Journal: Applied Economics, 8, 217246.
Singh M., Koran M. (2020), ’The virus doesn’t discriminate but governments do’:
Latinos disproportionately hit by coronavirus. The Guardian, April 18.
Tuttle W.M.Jr. (1970), Race Riot: Chicago in the Summer of 1919. New York:
Atheneum.
Voitglander N., Voth H. (2012), Persecution perpetuated: The medieval origins of
anti-Semitic violence in Nazi Germany. Quarterly Journal of Economics, 127, 1339–1392.
White C., Nafilyan V. (2020). Coronavirus (COVID-19) related deaths by ethnic
group, England and Wales: 2 March 2020 to 10 April 2020. Office for National Statistics.
Wiemers E.E., Abrahams S., AlFakhri M., Hotz V.J., Schoeni R.F., Seltzer J.A. (2020).
Disparities in vulnerability to severe complications from Covid-19 in the United States.
NBER Working Paper No. 27294.
Yancy C.W. (2020), COVID-19 and African Americans. Journal of the American
Medical Association. Opinion. April 15.
Zenou Y., Boccard N. (2000), Racial discrimination and redlining in cities. Journal of
Urban Economics, 48, 260–285.
170C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
APPENDIX
Figure A1: HOLC Area Description File for Area D96, Chicago Metropolitan Area, 1939
171C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
0.1
.2.3
Perc
ent o
f Dea
ths
Others Blacks
COVID-19 Deaths by Age Group
Vicenarian Tricenarian Quadragenarian Quinquagenarian Sexagenarian
Septuagenarian Octogenarian Nonagenariantransplant Centenarian
Figure A2: COVID-19 Deaths, by Age Group - Cook County, March 16-June 16, 2020
0.1
.2.3
.4.5
Perc
ent o
f Fem
ale
Dea
ths
Others Blacks
COVID-19 Deaths by Gender
Figure A3: COVID-19 Deaths, by Gender - Cook County, March 16-June 16, 2020
172C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
0.2
.4.6
Perc
ent o
f Dea
ths
Others Blacks
COVID-19 Deaths by Comorbidity
Diabetes Asthma Liver Cancer Hypertension
Kidney Obesity Respiratory Neuro-Cardiac Neuro-Respiratory
Asplenia Immunodeficiency Transplant Heart
Figure A4: COVID-19 Deaths, by Comorbidity - Cook County, March 16-June 16, 2020
0.1
.2.3
.4.5
Perc
ent o
f Bla
ck D
eath
s
A B C D
Percent of Black Deaths by HOLC Grade
Figure A5: Share of Black COVID-19 Deaths, by HOLC Grade - Cook County, March16-June 16, 2020
173C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
010
2030
40Pe
rcen
t of B
lack
s
A B C D
Percent of Blacks by HOLC Grade
Figure A6: Black Population Share, by HOLC Grade - Cook County, 2010 Census
Figure A7: HOLC Neighborhoods by Block GroupNote: The figure reports an example of four block groups, each partitioned into three neighborhoods, two of which be-longing to two different HOLC areas, both graded C (in yellow) and one ungraded (in white).
174C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
010
2030
40Pe
rcen
t of B
lock
Gro
ups
0 1 2 3 4 5 6 7 8 9 10HOLC Neighborhoods by Block Group
Figure A8: Distribution of HOLC Neighborhoods by Block GroupNote: The figure shows that over 40 percent of the block groups in the sample include only one HOLC neighborhood (po-tentially with the same HOLC grade), while around 35 percent include two, and so on, with a small percentage includingup to ten.
175C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
020
,000
40,0
0060
,000
80,0
00Pe
rson
al In
com
e
A B C D
05
1015
20Sh
are
With
No
Hig
h Sc
hool
A B C D
020
4060
8010
0Po
pula
tion
Shar
e
A B C D
Blacks Others
0.5
11.
52
Mor
talit
y R
ate
A B C D
Blacks Others
Figure A9: Socioeconomic and Demographic Characteristics by HOLC GradeNote: Personal income, share with no high school, population share by race, and mortality rate by race, by HOLC area.
176C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
0.0
5.1
.15
.20
.05
.1.1
5.2
w1 w2 w3 w4 w5 w6 w7 w8 w9w10w11w12w13w14w15w16w17w18w19w20w21w22w23w24 w1 w2 w3 w4 w5 w6 w7 w8 w9
w10w11w12w13w14w15w16w17w18w19w20w21w22w23w24
w1 w2 w3 w4 w5 w6 w7 w8 w9w10w11w12w13w14w15w16w17w18w19w20w21w22w23w24 w1 w2 w3 w4 w5 w6 w7 w8 w9
w10w11w12w13w14w15w16w17w18w19w20w21w22w23w24
A B
C D
Blacks Others
Mor
talit
y R
ate
Figure A10: Mortality Rate, by Race, Week, and HOLC Grade - Cook County, March16-June 16, 2020
177C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.05
0.0
5D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
a. Blacks
-.05
0.0
5D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 0 -1 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
b. Others
Figure A11: Including Block Groups Ungraded by the HOLCD - Cook County, January1-June 16, 2020Note: The dependent variable is number of deaths, of blacks (Panel a) and of other groups (Panel b). The sample includesblock groups that were not graded by the HOLC. The coefficients are least-squares estimates of the βks in regressions se-quentially including additional interactions between week dummies and the variables indicated for each panel. Block groupand week fixed effects are included. Vertical lines represent 95 percent confidence intervals based on standard errors clus-tered at at block group level. The omitted period k = −1, i.e., week 10.
178C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Age 65+
-.05
0.0
5.1
Blac
k D
eath
s-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13
Weeks
b. Chinese & Hispanic Share
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Unemployment & Income
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Education
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Black Share
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Distance from Hospital & Nursing Home
Figure A12: Dummy Death as Alternative Dependent Variable - Cook County, January1-June 16, 2020Note: The dependent variable is a dummy variable taking value one if a death, of blacks (Panel a) and of other groups(Panel b), occurred in a block group-week, and zero otherwise. The coefficients are least-squares estimates of the βks in re-gressions sequentially including additional interactions between week dummies and the variables indicated for each panel.Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals based on standarderrors clustered at at block group level. The omitted period k = −1, i.e., week 10.
179C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Age 65+
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. Chinese & Hispanic Share
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Unemployment & Income
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Education
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Black Share
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Distance from Hospital & Nursing Home
Figure A13: Allowing for Differential Trends - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks in re-gressions sequentially including additional interactions between week dummies and the variables indicated for each panel.Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals based on standarderrors clustered at at block group level. The omitted period k = −1, i.e., week 10.
180C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
Overlap One HOLC Neighborhood
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
Overlap up to Two HOLC Neighborhoods
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
Overlap up to Three HOLC Neighborhoods
-.05
0.0
5.1
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10+11+12+13Weeks
Overlap up to Four HOLC Neighborhoods
Figure A14: Robustness to Aggregation - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups including one, up to two, up to three, and up to four HOLC neighborhoods. Block group and weekfixed effects are included. Vertical lines represent 95 percent confidence intervals based on standard errors clustered at atblock group level. The omitted period k = −1, i.e., week 10.
181C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Personal Income Below Median
-.20
.2.4
Blac
k D
eath
s-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13
Weeks
b. Personal Income Above Median
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Personal Income Below Median & Black Share<25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Personal Income Below Median & Black Share>25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Personal Income Above Median & Black Share<25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Personal Income Above Median & Black Share>25%
Figure A15: Heterogeneity by Income - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with personal income below median (Panel a), above median (Panel b), below median and with blackshare below 25 percent (Panel c), below median and with black share above 25 percent (Panel d), above median and withblack share below 25 percent (Panel e), and above median and with black share above 25 percent (Panel f). Block groupand week fixed effects are included. Vertical lines represent 95 percent confidence intervals based on standard errors clus-tered at at block group level. The omitted period k = −1, i.e., week 10.
182C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Poverty Below Median
-.2-.1
0.1
.2.3
Blac
k D
eath
s-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13
Weeks
b. Poverty Above Median
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Poverty Below Median & Black Share<25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Poverty Below Median & Black Share>25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Poverty Above Median & Black Share<25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Poverty Above Median & Black Share>25%
Figure A16: Heterogeneity by Poverty - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with population share below the poverty line below median (Panel a), above median (Panel b), belowmedian and with black share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d),above median and with black share below 25 percent (Panel e), and above median and with black share above 25 percent(Panel f). Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals basedon standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
183C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Unemployment Below Median
-.20
.2.4
Blac
k D
eath
s-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13
Weeks
b. Unemployment Above Median
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Unemployment Below Median & Black Share<25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Unemployment Below Median & Black Share>25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Unemployment Above Median & Black Share<25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Unemployment Above Median & Black Share>25%
Figure A17: Heterogeneity by Unemployment - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with unemployment rate below median (Panel a), above median (Panel b), below median and withblack share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d), above median andwith black share below 25 percent (Panel e), and above median and with black share above 25 percent (Panel f). Blockgroup and week fixed effects are included. Vertical lines represent 95 percent confidence intervals based on standard errorsclustered at at block group level. The omitted period k = −1, i.e., week 10.
184C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. No High School Below Median
-.2-.1
0.1
.2.3
Blac
k D
eath
s-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13
Weeks
b. No High School Above Median
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. No High School Below Median & Black Share<25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. No High School Below Median & Black Share>25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. No High School Above Median & Black Share<25%
-.2-.1
0.1
.2.3
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. No High School Above Median & Black Share>25%
Figure A18: Heterogeneity by Education - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with population share with no high school below median (Panel a), above median (Panel b), belowmedian and with black share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d),above median and with black share below 25 percent (Panel e), and above median and with black share above 25 percent(Panel f). Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals basedon standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
185C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.20
.2.4
.6Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Age 65 Below Median
-.20
.2.4
.6Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
b. Age 65 Above Median
-.20
.2.4
.6Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Age 65 Below Median & Black Share<25%
-.20
.2.4
.6Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Age 65 Below Median & Black Share>25%
-.20
.2.4
.6Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Age 65 Above Median & Black Share<25%
-.20
.2.4
.6Bl
ack
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Age 65 Above Median & Black Share>25%
Figure A19: Heterogeneity by Age - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks over sam-ples of block groups with population share aged 65+ below median (Panel a), above median (Panel b), below median andwith black share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d), above me-dian and with black share below 25 percent (Panel e), and above median and with black share above 25 percent (Panel f).Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals based on standarderrors clustered at at block group level. The omitted period k = −1, i.e., week 10.
186C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
a. Single Parent Below Median
-.20
.2.4
Blac
k D
eath
s-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13
Weeks
b. Single Parent Above Median
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
c. Single Parent Below Median & Black Share<25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
d. Single Parent Below Median & Black Share>25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
e. Single Parent Above Median & Black Share<25%
-.20
.2.4
Blac
k D
eath
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
f. Single Parent Above Median & Black Share>25%
Figure A20: Heterogeneity by Single Parents - Cook County, January 1-June 16, 2020Note: The dependent variable is number of black deaths. The coefficients are least-squares estimates of the βks oversamples of block groups with population share of single parents below median (Panel a), above median (Panel b), belowmedian and with black share below 25 percent (Panel c), below median and with black share above 25 percent (Panel d),above median and with black share below 25 percent (Panel e), and above median and with black share above 25 percent(Panel f). Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals basedon standard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
187C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
010
2030
40
18mar2020 08apr2020 29apr2020 20may2020 10jun2020Week
Hispanics OthersBlacks
Figure A21: Hispanic, Black, and Other COVID-19 Deaths - Cook County, March 16-June 16, 2020Note: The figure reports the number of COVID-19 related deaths by day, separately for Hispanics, blacks, and othergroups.
188C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
0.0
5.1
.15
.2M
orta
lity
Rat
e
2020w1 2020w5 2020w9 2020w14 2020w18 2020w22Week
Others Hispanics Blacks
Figure A22: Mortality Rate, Hispanics, Blacks, and Others - Cook County, January1-June 16, 2020Note: Mortality rates from any cause of death by week, separately for Hispanics, blacks, and other groups.
189C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
-.02
0.0
2.0
4.0
6H
ispa
nic
Dea
ths
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +10 +11 +12 +13Weeks
Hispanics
Figure A23: Dynamic Effect of the Treatment on Number of Hispanic Deaths - CookCounty, January 1-June 16, 2020Note: The dependent variable is number of Hispanic deaths. The coefficients are least-squares estimates of the βks with10k9. Block group and week fixed effects are included. Vertical lines represent 95 percent confidence intervals based onstandard errors clustered at at block group level. The omitted period k = −1, i.e., week 10.
190C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A1: Variable Definitions and Sources - Cross Section
Variable Definition Source
Black COVID-19 Death Dummy variable taking value one if an individ-ual who died from COVID-19 is black, and zerootherwise
Cook County Medical Examiner’s Officer
Hispanic COVID-19Death
Dummy variable taking value one if an individualwho died from COVID-19 is Hispanic, and zerootherwise
Cook County Medical Examiner’s Officer
Other COVID-19 Death Dummy variable taking value one if an individualwho died from COVID-19 is other than black, andzero otherwise
Cook County Medical Examiner’s Officer
HOLC A Dummy variable taking value one if an individualwho died from COVID-19 lived in HOLC Area A,and zero otherwise
University of Richmond Mapping Inequality:Redlining in New Deal America 1935-1940
HOLC B Dummy variable taking value one if an individualwho died from COVID-19 lived in HOLC Area B,and zero otherwise
University of Richmond Mapping Inequality:Redlining in New Deal America 1935-1940
HOLC C Dummy variable taking value one if an individualwho died from COVID-19 lived in HOLC Area C,and zero otherwise
University of Richmond Mapping Inequality:Redlining in New Deal America 1935-1940
HOLC D Dummy variable taking value one if an individualwho died from COVID-19 lived in HOLC Area D,and zero otherwise
University of Richmond Mapping Inequality:Redlining in New Deal America 1935-1940
Age Groups Set of eight dummy variables taking value oneif an individual who died from COVID-19 wasrespectively aged from 20-29 (tricenarian) up to100+ (centenarian), and zero otherwise
Cook County Medical Examiner’s Officer
Female Dummy variable taking value one if an individualwho died from COVID-19 was female
Cook County Medical Examiner’s Officer
Comorbidities Set of 14 dummy variables that take value one(and zero otherwise) when an individual who diedfrom COVID-19 was respectively affected by dia-betes and/or asthma, liver disease, cancer, hyper-tension, kidney disease, obesity, respiratory dis-eases (including cystic fibrosis, pulmonary andlung diseases), neuro-cardiac diseases (includ-ing cardiovascular disease, stroke, and demen-tia), neuro-respiratory diseases (including scle-rosis, Parkinson, myastenia, palsy, hemiplegia,quadriplegia, brain and cerebellum diseases), as-plenia (including spenectomy, spleen and sicklecell disease), immunodeficiency (including HIV,immunosuppression, and anaemia), transplant,and heart diseases (including valve disease).
Cook County Medical Examiner’s Officer
Distance From Hospital Individual distance (in degrees) from hospital Cook County Health and Hospitals FacilitiesDistance From NursingHome
Individual distance (in degrees) from nursinghome
Medicare Nursing Home Compare
Population 2014-2018 tract level population American Community Survey through CookCounty Government Open Data
Share Aged 18-64 2014-2018 tract level share of the population aged18-64
American Community Survey through CookCounty Government Open Data
Share Aged 65+ 2014-2018 tract level share of the population aged65+
American Community Survey through CookCounty Government Open Data
Share With No HighSchool
2014-2018 tract level share of the populationwithout a high school diploma
American Community Survey through CookCounty Government Open Data
Black Share 2014-2018 tract level share of the population whois black
American Community Survey through CookCounty Government Open Data
Hispanic Share 2014-2018 tract level share of the population whois Hispanic
American Community Survey through CookCounty Government Open Data
191C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A2: Variable Definitions and Sources - Panel
Variable Definition Source
Deaths Number of deaths from any cause Cook County Medical Examiner’s OfficerBlack Deaths Number of deaths from any cause of black indi-
vidualsCook County Medical Examiner’s Officer
Other Deaths Number of deaths from any cause of other thanblack individuals
Cook County Medical Examiner’s Officer
Hispanic Deaths Number of deaths from any cause of Hispanic in-dividuals
Cook County Medical Examiner’s Officer
Death Dummy Dummy variable taking value one if a death oc-curred in a block group-week, and zero otherwise
Cook County Medical Examiner’s Officer
Black Death Dummy Dummy variable taking value one if a black deathoccurred in a block group-week, and zero other-wise
Cook County Medical Examiner’s Officer
Other Dummy Dummy variable taking value one if an other thanblack death occurred in a block group-week, andzero otherwise
Cook County Medical Examiner’s Officer
Black Mortality Rate Number of deaths from any cause of black indi-viduals over tract level black population
Cook County Medical Examiner’s Officer andAmerican Community Survey through CookCounty Government Open Data
Other Mortality Rate Number of deaths from any cause of other thanblack individuals over tract level other than blackpopulation
Cook County Medical Examiner’s Officer andAmerican Community Survey through CookCounty Government Open Data
Hispanic Mortality Rate Number of deaths from any cause of Hispanic in-dividuals over tract level Hispanic population
Cook County Medical Examiner’s Officer andAmerican Community Survey through CookCounty Government Open Data
Treated Groups Majority C & D block groups University of Richmond Mapping Inequality:Redlining in New Deal America 1935-1940
Share Aged 65+ 2014-2018 tract level share of the population aged65+
American Community Survey through CDC So-cial Vulnerability Index
Hispanic Share 2014-2018 tract level share of the population whois Hispanic
American Community Survey through CDC So-cial Vulnerability Index
Share Chinese PrimaryLanguage
2014-2018 tract level share of the populationspeaking Chinese as primary language
American Community Survey through CDC So-cial Vulnerability Index
Personal Income 2014-2018 tract level personal income American Community Survey through CDC So-cial Vulnerability Index
Unemployment Rate 2014-2018 tract level unemployment rate American Community Survey through CDC So-cial Vulnerability Index
Share With No HighSchool
2014-2018 tract level share of the populationwithout a high school diploma
American Community Survey through CDC So-cial Vulnerability Index
Black Share 2014-2018 tract level share of the population whois black
American Community Survey through CDC So-cial Vulnerability Index
Distance From Hospital Minimal centroid distance of a block group fromhospital
Cook County Health and Hospitals Facilities
Distance From NursingHome
Minimal centroid distance of a block group fromnursing home
Medicare Nursing Home Compare
Socioeconomic StatusVulnerability
Index comprising personal income, share belowpoverty line, unemployment rate, and share with-out high school
CDC Social Vulnerability Index
Household CompositionVulnerability
Index comprising share aged 65+, share aged 17-, share aged 5+with disability, share of single-parent households
CDC Social Vulnerability Index
Minority Status Vulner-ability
Index comprising population share in minoritystatus and “Speak English Less than Well”
CDC Social Vulnerability Index
Housing Vulnerability Index comprising housing share of multi-housingstructures, mobile homes, and more people perroom and population share with no vehicle avail-able and in group quarters
CDC Social Vulnerability Index
Overall Vulnerability Index based on the above four indices CDC Social Vulnerability IndexShare Below Poverty 2014-2018 tract level share of the population be-
low poverty lineAmerican Community Survey through CDC So-cial Vulnerability Index
Share Single Parents 2014-2018 tract level share of the population insingle-parent households
American Community Survey through CDC So-cial Vulnerability Index
192C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A3: Summary Statistics, COVID-19 Deaths, Cross Section - Cook County, March16-June 16, 2020
count mean sd min maxBlack COVID-19 Death 3621 0.353 0.478 0.000 1.000Hispanic COVID-19 Death 3621 0.186 0.389 0.000 1.000Other COVID-19 Death 3621 0.647 0.478 0.000 1.000HOLC A 3621 0.002 0.047 0.000 1.000HOLC B 3621 0.076 0.265 0.000 1.000HOLC C 3621 0.381 0.486 0.000 1.000HOLC D 3621 0.141 0.348 0.000 1.000No HOLC 3621 0.400 0.490 0.000 1.000Age Groups 3619 5.912 1.509 1.000 9.000Female 3621 0.423 0.494 0.000 1.000Diabetes 3621 0.406 0.491 0.000 1.000Asthma 3621 0.042 0.200 0.000 1.000Liver Diseases 3621 0.006 0.076 0.000 1.000Cancer 3621 0.029 0.169 0.000 1.000Hypertension 3621 0.524 0.499 0.000 1.000Kidney Disease 3621 0.107 0.309 0.000 1.000Obesity 3621 0.087 0.282 0.000 1.000Respiratory Diseases 3621 0.157 0.364 0.000 1.000Neuro-cardiac Diseases 3621 0.257 0.437 0.000 1.000Neuro-respiratory Diseases 3621 0.035 0.184 0.000 1.000Asplenia 3621 0.001 0.023 0.000 1.000Immunodeficiency 3621 0.006 0.076 0.000 1.000Transplant 3621 0.005 0.070 0.000 1.000Heart Diseases 3621 0.113 0.317 0.000 1.000Population 3621 4666.875 1776.020 592.000 19015.000Share Aged 18-64 3621 62.219 7.527 38.300 94.000Share Aged 65+ 3621 15.557 8.022 0.000 51.000Distance From Hospital 3621 0.023 0.016 0.000 0.116Distance From Nursing Home 3621 0.011 0.012 0.000 0.067Share With No High School 3621 17.061 11.926 0.000 60.600Black Share 3621 30.940 36.592 0.000 100.000Hispanic Share 3621 25.903 28.420 0.000 99.600
193C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A4: Black COVID-19 Death, Cross-Sectional Results, All Coefficients - CookCounty, March 16-June 16, 2020
Black COVID-19 Death(1) (2) (3) (4) (5) (6) (7) (8)
HOLC A -0.3808*** -0.3652*** -0.3244*** -0.3261*** -0.3247*** -0.3636*** -0.3514*** -0.1250***(0.0640) (0.0720) (0.0757) (0.0800) (0.0811) (0.0798) (0.0751) (0.0424)
HOLC B 0.0389 0.0360 0.0399 0.0373 0.0355 0.0303 0.0072 0.0169(0.0329) (0.0334) (0.0327) (0.0323) (0.0327) (0.0322) (0.0316) (0.0250)
HOLC C 0.0800*** 0.0817*** 0.0830*** 0.0677*** 0.0716*** 0.1176*** 0.0848*** -0.0146(0.0177) (0.0185) (0.0181) (0.0175) (0.0202) (0.0216) (0.0234) (0.0182)
HOLC D 0.1752*** 0.1769*** 0.1745*** 0.0833*** 0.0852*** 0.1232*** 0.0922*** 0.0106(0.0233) (0.0238) (0.0231) (0.0255) (0.0266) (0.0281) (0.0297) (0.0222)
Tricenarian 0.1373 0.1069 0.0898 0.0934 0.1113 0.1244 -0.0055(0.1324) (0.1333) (0.1306) (0.1318) (0.1287) (0.1278) (0.0981)
Quadragenarian 0.0593 0.0282 0.0278 0.0326 0.0600 0.0728 0.0233(0.1262) (0.1278) (0.1282) (0.1294) (0.1255) (0.1236) (0.0961)
Quinquagenarian 0.0559 0.0140 0.0198 0.0252 0.0480 0.0648 0.0121(0.1225) (0.1229) (0.1230) (0.1240) (0.1211) (0.1192) (0.0895)
Sexagenarian 0.1348 0.0778 0.0795 0.0857 0.0907 0.1144 0.0603(0.1200) (0.1216) (0.1207) (0.1220) (0.1197) (0.1178) (0.0911)
Septuagenarian 0.1149 0.0543 0.0619 0.0678 0.0672 0.0890 0.0288(0.1171) (0.1178) (0.1177) (0.1189) (0.1168) (0.1145) (0.0892)
Octogenarian 0.0768 0.0148 0.0155 0.0218 0.0140 0.0409 0.0180(0.1188) (0.1192) (0.1200) (0.1212) (0.1196) (0.1182) (0.0891)
Nonagenarian 0.0082 -0.0498 -0.0433 -0.0374 -0.0468 -0.0102 -0.0291(0.1226) (0.1225) (0.1233) (0.1248) (0.1230) (0.1203) (0.0921)
Centenarian 0.0986 0.0372 0.0399 0.0445 0.0270 0.0744 0.0061(0.1350) (0.1358) (0.1337) (0.1356) (0.1338) (0.1311) (0.1046)
Female 0.1036*** 0.0977*** 0.0961*** 0.0955*** 0.0886*** 0.0902*** 0.0504***(0.0156) (0.0157) (0.0154) (0.0154) (0.0155) (0.0155) (0.0115)
Diabetes 0.0040 -0.0004 -0.0003 0.0103 0.0092 0.0005(0.0156) (0.0155) (0.0154) (0.0152) (0.0149) (0.0122)
Asthma 0.0145 0.0155 0.0147 0.0085 -0.0003 -0.0034(0.0440) (0.0438) (0.0438) (0.0427) (0.0427) (0.0314)
Liver Diseases -0.1527 -0.1488* -0.1498* -0.1516* -0.1496* 0.0089(0.0948) (0.0870) (0.0868) (0.0820) (0.0801) (0.0771)
Cancer 0.0722 0.0747 0.0753 0.0670 0.0580 0.0166(0.0478) (0.0467) (0.0474) (0.0480) (0.0477) (0.0395)
Hypertension 0.0939*** 0.0936*** 0.0938*** 0.0908*** 0.0881*** 0.0545***(0.0165) (0.0166) (0.0166) (0.0167) (0.0167) (0.0142)
Kidney Disease 0.0782*** 0.0806*** 0.0802*** 0.0694** 0.0702** 0.0659***(0.0272) (0.0266) (0.0268) (0.0266) (0.0269) (0.0213)
Obesity 0.0413 0.0305 0.0306 0.0290 0.0264 0.0118(0.0297) (0.0287) (0.0289) (0.0281) (0.0280) (0.0234)
Respiratory Diseases 0.0764*** 0.0740*** 0.0747*** 0.0700*** 0.0752*** 0.0405**(0.0214) (0.0209) (0.0209) (0.0203) (0.0201) (0.0175)
Neuro-cardiac Diseases 0.0158 0.0166 0.0177 0.0196 0.0236 0.0453***(0.0213) (0.0214) (0.0217) (0.0217) (0.0213) (0.0165)
Neuro-respiratory Diseases -0.0997*** -0.0951*** -0.0943*** -0.0986*** -0.0843** -0.0739***(0.0338) (0.0333) (0.0330) (0.0327) (0.0338) (0.0261)
Asplenia 0.1501 0.1260 0.1394 0.1531 0.1704 -0.1478(0.3863) (0.3213) (0.3237) (0.3323) (0.3184) (0.1288)
Immunodeficiency 0.4063*** 0.3897*** 0.3927*** 0.3756*** 0.3685*** 0.2470**(0.0899) (0.0934) (0.0935) (0.0883) (0.0877) (0.1020)
Transplant 0.1204 0.1200 0.1177 0.0940 0.0805 0.0136(0.1186) (0.1159) (0.1166) (0.1071) (0.1071) (0.1024)
Heart Diseases 0.0256 0.0178 0.0180 0.0093 0.0157 0.0081(0.0267) (0.0271) (0.0271) (0.0266) (0.0259) (0.0197)
Population (log) -0.1734*** -0.1725*** -0.1780*** -0.1568*** -0.0260(0.0210) (0.0208) (0.0209) (0.0207) (0.0178)
Share Aged 18-64 -0.0016 -0.0065*** -0.0067*** 0.0014(0.0014) (0.0018) (0.0017) (0.0012)
Share Aged 65+ -0.0007 -0.0065*** -0.0058*** -0.0011(0.0013) (0.0015) (0.0016) (0.0011)
Share With No High School -0.0073*** -0.0073*** -0.0007(0.0009) (0.0008) (0.0006)
Distance From Hospital -2.9815*** 0.4574(0.4945) (0.4217)
Distance From Nursing Home 3.5242*** -1.5140***(0.6538) (0.5396)
Black Share 0.0086***(0.0002)
Constant 0.2957*** 0.1637 0.1384 1.6084*** 1.6999*** 2.2492*** 2.0984*** 0.1585(0.0094) (0.1185) (0.1233) (0.2004) (0.2262) (0.2516) (0.2511) (0.2039)
Adj.R-squared 0.056 0.069 0.090 0.111 0.111 0.132 0.144 0.475Observations 3618 3616 3616 3616 3616 3616 3616 3616
Note: The dependent variable is a dummy variable that takes value one if an individual who died from COVID-19 is black,an zero otherwise. The omitted HOLC area is A. The omitted age group is vicenarian. Robust standard errors clusteredat a day level in parentheses: *** p<0.01, ** p<0.05, * p<0.1.
194C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A5: Summary Statistics, All Deaths, Panel - Cook County, January 1-June 16,2020
count mean sd min maxDeaths 95808 0.070 0.342 0.000 13.000Black Deaths 95808 0.029 0.201 0.000 10.000Other Deaths 95808 0.042 0.257 0.000 11.000Hispanic Deaths 95808 0.012 0.117 0.000 7.000Death Dummy 95808 0.057 0.233 0.000 1.000Black Death Dummy 95808 0.025 0.156 0.000 1.000Other Death Dummy 95808 0.035 0.184 0.000 1.000Black Mortality Rate 93576 0.064 3.845 0.000 1000.000Other Mortality Rate 95544 0.030 0.806 0.000 125.000Hispanic Mortality Rate 93504 0.029 3.324 0.000 1000.000White Mortality Rate 94961 0.055 1.909 0.000 333.333Treated Groups (Majority C & D) 95808 0.619 0.486 0.000 1.000Post-treatment (Week After 11) 95808 0.583 0.493 0.000 1.000Treatment 95808 0.361 0.480 0.000 1.000Share Aged 65+ 95736 14.156 6.262 0.000 51.300Hispanic Share 95808 23.627 25.999 0.000 99.600Share Chinese Primary Language 95808 1.113 4.582 0.000 78.000Personal Income 95736 34845.017 20754.386 2530.000 154760.000Unemployment Rate 95808 9.930 7.949 0.000 92.820Share With No High School 95736 13.847 10.704 0.000 61.600Black Share 95808 27.330 35.880 0.000 100.000Distance From Hospital 95808 0.021 0.016 0.000 0.105Distance From Nursing Home 95808 0.013 0.011 0.000 0.066Socioeconomic Status Vulnerability 95736 0.539 0.303 0.000 1.000Household Composition Vulnerability 95736 0.454 0.286 0.002 0.999Minority Status Vulnerability 95736 0.674 0.221 0.035 1.000Housing Vulnerability 95736 0.544 0.270 0.000 1.000Overall Vulnerability 95736 0.565 0.284 0.001 0.998Share Below Poverty 95736 15.883 12.047 0.300 77.100Share Single Parents 95808 9.896 7.413 0.000 51.300
Table A6: Average Treatment Effect
(1) (2)Black Deaths Other Deaths
Treatment 0.015** -0.002(0.006) (0.009)
R-squared 0.090 0.082Observations 53903 53903Mean Untreated In Post-treatment 0.025 0.025
Note: The dependent variable is the average number of deaths in a block group in a week. Robust standard errors clus-tered at a block groups level in parentheses: *** p<0.01, ** p<0.05, * p<0.1.
Table A7: Summary Statistics, Spanish Flu Deaths, Chicago, 1918
count mean sd min maxDeaths From Spanish Flu 3472 2.296 2.974 0.000 31.000Majority C & D 3419 0.848 0.359 0.000 1.000Population Density (log) 3472 3.488 1.098 -3.640 5.205Illiteracy Rate 3472 0.040 0.048 0.000 0.279Home Ownership Rate 3472 0.061 0.038 0.000 0.197Unemployment Rate 3472 0.346 0.050 0.058 0.495
195C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
129
-195
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Covid Economics Issue 38, 16 July 2020
Copyright: Michèle Belot, Syngjoo Choi, Egon Tripodi, Eline van den Broek-Altenburg, Julian C. Jamison and Nicholas W. Papageorge
Unequal consequences of Covid-19 across age and income: Representative evidence from six countries1
Michèle Belot,2 Syngjoo Choi,3 Egon Tripodi,4 Eline van den Broek-Altenburg,5 Julian C. Jamison6 and Nicholas W. Papageorge7
Date submitted: 14 July 2020; Date accepted: 14 July 2020
Covid-19 and the measures taken to contain it have led to unprecedented constraints on work and leisure activities, across the world. This paper uses nationally representative surveys to document how people of different ages and incomes have been affected across six countries (China, South Korea, Japan, Italy, UK and US). We first document changes in economic variables (income and consumption) and leisure. Second, we document self-reported negative and positive non-financial effects of the crisis. We then examine attitudes towards recommendations (wearing a mask in particular) and the approach taken by public authorities. We find similarities across countries in how people of different generations have been affected. Young people have experienced more drastic changes to their lives, and overall they are less supportive of these measures. These patterns are less clear across income groups: while some countries have managed to shield lower income individuals from negative consequences, others have not. We also show that how people have been affected by the crisis (positively or negatively) does little to explain whether or not they support measures implemented by the public authorities. Young people are overall less supportive of such measures independently of how they have been affected.
1 Survey and data collection protocol were approved by the ethics board at the University of Exeter (application id eUEBS003014v2.0). Research funding from the Creative-Pioneering Researchers Program at Seoul National University, and from the European University Institute are gratefully acknowledged. This paper uses data that is publicly available at https://osf.io/aubkc/.
2 Professor of Economics, European University Institute.3 Professor of Economics, Seoul National University.4 PhD Candidate, European University Institute.5 Assistant Professor, University of Vermont.6 Professor of Economics, University of Exeter.7 Broadus Mitchell Associate Professor of Economics, Johns Hopkins University.
196C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
1 Introduction
The Covid-19 pandemic has affected almost all countries in the world and has led to un-precedented measures being implemented to contain the virus. Countries have differed intheir response to the epidemic. Some adopted stringent measures, such as shelter-in-placeorder, while others implemented early and widespread testing and tracing procedures.
The adjustments required to contain the epidemic have had a dramatic impact onhow we live, on our ability to work and on our leisure activities. A key concern is thatthe groups that have been affected most by the measures taken are not the ones whoface the highest risks of severe illness. Such misalignment between personal incentivesand burdens on the one hand, and public health concerns on the other hand, are a mainchallenge in devising and implementing effective public policies.
Without evidence that improves our understanding on the nature of such misalign-ment, our ability to contain the epidemic and reduce economic and social damages is lim-ited. This paper offers survey-based evidence from six countries on the heterogeneousnature of the economic and social consequences of Covid-19 along with information onbehavioral response to the crisis and attitudes towards government measures that havebeen implemented. Our work complements preliminary evidence put forward in a fewrecent studies focusing on specific countries and specific aspects (economic) of the crisissuch as Adams et al. (2020), Montenovo et al. (2020), Fairlie et al. (2020), von Gaudeckeret al. (2020).
We document how the experience of the epidemic and measures that have been im-plemented have differed according to two key individual characteristics: age and income.The evidence is based on data collected in the third week of April 2020 on samples ofaround 6,000 individuals from three Western countries—US, UK and Italy—and threeAsian countries—China, Japan and South Korea (Belot et al., 2020). The samples are na-tionally representative on three dimensions: age, gender and income. We focus in thispaper on how age and income gradients relate to the Covid pandemic.
At the time of data collection, countries we examined were at different phases of theepidemic and had implemented different measures.1 These differences, on top of differ-ences in other factors (such as cultural attitudes), can all contribute to explain the cross-country differences in the nature of the Covid-19 effects. Instead of identifying the causesof such differences, this paper marks a first step in understanding how the pandemic hasaffected different age and income groups across countries.
First, we document objective changes in two key aspects of life: (1) economic variables(income/consumption); and (2) leisure and social life. In particular, we examine whetherpeople experienced a loss in income and a drop in spending during the pandemic. Next,we look at how people reduced the frequency of different behaviors that have a social
1See for the six countries, in Figures A1 and A2 of the Online Supplementary Material, how the datacollection time window overlaps with the time series of the number of confirmed infected people anddeaths per million inhabitants together with the stringency index of government policies to contain thepandemic, based on the Oxford Covid-19 Government Response Tracker data (Hale et al., 2020).
197C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
component—attending large social gatherings, visiting family and friends, and going tolarge close or open public spaces.
Second, we examine subjective non-financial consequences that individuals reportexperiencing. Negative effects we consider include boredom, loneliness, trouble sleep-ing, anxiety and stress, and conflicts with friends and neighbors. These negative, non-financial effects are potentially important because they speak to the burden of complyingwith measures to contain the pandemic and also shape incentives for individuals to fol-low social distance measures. Positive effects we examine include spending more timewith family, enjoying more free time, and reductions in pollution and noise.
Third, we look at a measure of a specific behavior that has aspects of solidarity andprecaution, in particular, wearing a mask. In the presence of strong externality, the bene-fits of wearing a mask are understood to be not only to protect oneself but also to reducechances of transmitting the corona virus to other people (Chu et al., 2020; Howard et al.,2020).
Finally, we examine measures of support for the approach taken by each country’sgovernment, and examine to what extent differences in support can be explained by dif-ferences in the impact of the pandemic on individuals.
2 Materials and Methods
2.1 Data
We use publicly available data that was collected by Belot et al. (2020) between April15 and April 23. This dataset includes 6,082 respondents; roughly 1000 from each ofsix countries. Three Asian countries (China, Japan and South Korea) and three westerncountries (Italy, the United Kingdom and the United States). For each country, the sampleis nationally representative along age, gender, and (pre-Covid) household income. In theUnited States the data includes respondents from the 4 most populous states: California,Florida, New York and Texas. American respondents self-identify their race, and thesample is also nationally representative along this dimension.
2.2 Methods
Our analysis is based on ordinary least square models (or linear probability models whenthe outcome is binary). The right hand side variables include age and income dummies,as well as additional control variables such as gender, a rural-urban indicator, and re-gional dummies. The age categories we consider are: below 25 (between 18 and 25),26-45, 46-65 and above 65. For income, we use the categorical variable indicating thehousehold income quintile as reported by the respondent.
We first examine the extent to which groups of different ages and income quintileshave been affected differently in their economic situation and in their social life. To assess
198C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
the economic impact, we consider two key variables: (1) Experience of a fall in householdincome and (2) Experience of a drop in spending.
To assess the impact on social life and leisure, we construct an index measuring thedegree of engagement in different leisure activities that have a social component. Respon-dents were asked to indicate the frequency of engagement in a series of activities at dif-ferent points in time: in normal times before the outbreak, at the start of the outbreak andat the time of the survey. The index aggregates information from four variables: partici-pation in large social gatherings, going to large close spaces (such as a museum or a shop-ping center), going to large open spaces (such as a public park), visiting friends/family.
Second, we examine self-reported positive and negative non-financial effects of thepandemic and measures implemented. Negative effects include anxiety, trouble sleeping,increased conflicts, boredom or loneliness. Positive effects include more time with family,more free time, less pollution or less noise. Survey participants could indicate as manyas applicable. We construct two simple indicators of the number of positive and negativeeffects indicated.
Third, we look at one specific behavior that appears to exhibit large cross-countryvariation: wearing a facial mask. This behavior is interesting because it is not very costly,and it has a clear element of solidarity, since the main benefit appears to reduce transmis-sion to others (rather than protecting oneself). However, countries did not universallyrecommended the use of masks in their population, at least not early on.
The final variable of interest is the support for the approach taken by the country’sgovernment. Here we will highlight age and income differences in beliefs of effective-ness of measures implemented and in the general support for the approach taken by thegovernment. We explore to what extent these can be explained by the variables capturingthe economic and social impact of the pandemic. We present an analysis where we addcontrols for variables capturing the economic and social impact as described above.
Note that we will interpret significance levels of the coefficients at face value, withoutimplementing corrections for multiple hypotheses testing. Given how little is knownon the topic this analysis is necessarily exploratory, and confirmatory research will beneeded. Yet, the goal is to see if a coherent story emerges.
3 Results
3.1 Economic and social consequences of the pandemic
We find evidence of a negative age gradient in the probability of having experienced afall in household income across all countries except South Korea, as shown in Figure 1.The oldest group (65+) is 47 percentage points less likely to have experienced a drop inincome in China, relative to the youngest group (18-25), the difference is large but lesspronounced in other countries (around 25 percentage points in Japan, Italy and the US,and 35 percentage points in the UK). For income, we find a less clear pattern, except for
199C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Italy and Korea, where those with incomes in the top 20% are significantly less likely(by 13 and 16 percentage points, respectively) to have experienced a fall in householdincome.2
Note that in all countries, we see a very similar pattern in the probability of telework-ing (see Table A3): Younger groups and higher income groups are substantially morelikely to be teleworking than those in the bottom 20% income. In China, the 46-65 are 25percentage points less likely to telework relative to to the 18-25 group. The difference issmaller but remains large in other countries, except for the US and Italy, where there is nosignificant difference.
Regarding spending, we see a similar pattern as with income loss, according to age.Older groups are less likely to have experienced a fall in consumption. We also see apositive income gradient, but only in the UK and the US: The higher income groups aremore likely to have experienced a drop in consumption.3 The most likely explanation isthe closure of shops and leisure-related facilities.
The picture that emerges is heterogeneous: Some countries appear to have succeededin the early months of the pandemic to shield lower income groups from negative finan-cial effects (like China, Japan, UK and US), while others did not (South Korea, Italy).
(a) Experienced income loss
18-25
26-45
46-65
Above 65
Income Q1
Income Q2
Income Q3
Income Q4
Income Q5
-.6 -.4 -.2 0 .2 -.6 -.4 -.2 0 .2 -.6 -.4 -.2 0 .2 -.6 -.4 -.2 0 .2 -.6 -.4 -.2 0 .2 -.6 -.4 -.2 0 .2
China Japan Korea Italy UK US
2We do not present cross-country comparisons of job loss because the financial implications of job lossvary across countries, depending on transfer programs that have been implemented as a result of the crisis.For the US, Papageorge et al. (2020) show an income gradient in the probability of permanent job loss.
3This echoes evidence from credit card transaction data reported in Chetty et al. (2020) for the US.
200C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
(b) Experienced drop in spending
18-25
26-45
46-65
Above 65
Income Q1
Income Q2
Income Q3
Income Q4
Income Q5
-.4 -.2 0 .2 .4 -.4 -.2 0 .2 .4 -.4 -.2 0 .2 .4 -.4 -.2 0 .2 .4 -.4 -.2 0 .2 .4 -.4 -.2 0 .2 .4
China Japan Korea Italy UK US
Country average for: China Japan Korea Italy UK USExperienced income loss (share) 0.603 0.385 0.478 0.579 0.428 0.486Experienced drop in spending (share) 0.565 0.308 0.489 0.566 0.396 0.431
Note: Point estimates and 95% confidence intervals from a linear probability model of an indicator variable. This indicator denotesloss of household income during the pandemic in panel (a), drop in consumption in panel (b). Covariates include income quintile,age group, gender and geographical controls. Figure based on regression results in Tables A1, A2. 18-25 and Income Q1 are baselinecategories for age and income quintile groups, respectively. Country averages of the outcome variables used for the regression arereported below the figure.
Figure 1: Age and income gradients on drop in household income and spending
Turning to social interactions, Figure 2 shows the mean reported levels of our indexvariable at three points in time—in normal times before the outbreak, at the start of theoutbreak and at the time of the survey. In all countries, the younger groups (18-25 or26-45) are most engaged in social activities. But the older groups appear to have reducedtheir social life most. There is also a clear income gradient: Higher income groups aremore likely to engage in leisure activities with a social component, in all countries. Sincethose were effectively discouraged or forbidden at the time of the survey, higher incomegroups by then had experienced a larger negative impact on their social life in most of thecountries. This is evidenced by marked income gradients on how bothered they reportbeing for not being able to participate in large social gatherings, go to large (close or open)spaces, and visit friends or family (see Table A4).
201C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
18-2526-4546-65
Above 66Income Q1Income Q2Income Q3Income Q4Income Q5
NeverRarely
Sometimes
Very often
Always
NeverRarely
Sometimes
Very often
Always
NeverRarely
Sometimes
Very often
Always
NeverRarely
Sometimes
Very often
Always
NeverRarely
Sometimes
Very often
Always
NeverRarely
Sometimes
Very often
Always
China Japan Korea Italy UK US
Normal times Start pandemic Time of survey
Note: We report group averages of an index that includes frequency of (i) participation in large social gatherings, (ii) visit to large openspaces, (iii) large close spaces, and (iv) visits to friends or family. The index is constructed by averaging (i)-(iv) frequencies on a 1 to 5scale, where ”1” is ”Never” and ”5” is ”Always”.
Figure 2: Social interactions over time, by age and income groups
Summarizing and looking across countries, we find that those who experienced thelargest negative economic impacts are the young, while older groups and high incomegroups experienced the largest negative impact in their social life and leisure.
3.2 Psychological costs and the positive side of the pandemic
Looking at negative non-financial effects, we find that the younger groups are more likelyto report negative effects, in all countries. Understanding the higher psychological costsof the younger groups is important because they may comply less with social distancingmeasures. Again, the pattern is less clear across income groups: There is no gradient inChina, Korea, Italy and the US, but there is a negative income gradient in the UK and apositive one in Japan.
We also find that people report experiencing some benefits from the pandemic—betweenenjoying more free time, enjoying time with family, cleaner air, and less noise pollution.The older groups are less likely to report positive effects (Table A7). We see a positiveincome gradient in Japan, Italy, and the US, where people in the lowest income quintilereport fewer positive effects from the crisis.
Summarizing, we find that young people are most affected (negatively and positively)in non-financial, psychological terms; all income groups appear to experience negativeeffects, but positive effects appear concentrated among the higher income groups.
202C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
18-25
26-45
46-65
Above 65
Income Q1
Income Q2
Income Q3
Income Q4
Income Q5
-1.5 -1 -.5 0 .5 -1.5 -1 -.5 0 .5 -1.5 -1 -.5 0 .5 -1.5 -1 -.5 0 .5 -1.5 -1 -.5 0 .5 -1.5 -1 -.5 0 .5
China Japan Korea Italy UK US
China Japan Korea Italy UK USCountry average 1.675 1.228 1.408 1.640 1.659 1.693
Note: Point estimates and 95% confidence intervals from a linear regression model of number of negative non-financial effects dueto the pandemic (which include: (i) boredom, (ii) loneliness, (iii) trouble sleeping, (iv) general anxiety and stress, and (v) increasedconflicts with friends/family/neighbors) on income quintile, age group, gender and geographical controls. Figure based on regressionresults in Table A5. 18-25 and Income Q1 are baseline categories for age and income quintile groups, respectively. Country averagesof the outcome variable used for the regression are reported below the figure.
Figure 3: Age and income gradients on negative non-financial effects
3.3 Wearing a facial mask
We now look at the age and income gradients in the probability of wearing a mask Fig-ure 4. This behavior is interesting because it involves a relatively low cost and it has aclear solidarity component, since the benefit appears to accrue mostly to others ratherthan oneself (Chu et al., 2020; Howard et al., 2020).
Here we find striking differences across countries. First, on average, the willingnessto wear a mask is much higher in Asian countries and Italy than in the UK and the US.Second, in Asian countries, there is hardly any gradient along age or income. In Korea,only the bottom 20% income group appears more reluctant to wear a mask. In the US,there is a clear positive income gradient.4 In the UK differences are most pronouncedaccording to age, with the older groups much more reluctant to wear a mask. In Italy, theage gradient goes in the opposite direction.
4In a related paper focusing on the US and using the same data (Papageorge et al. (2020)), we studychanges in a wider range of self-protective behaviors in the United States. We find that higher incomegroups are more likely to adopt self-protective measures in response to the outbreak.
203C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
18-25
26-45
46-65
Above 65
Income Q1
Income Q2
Income Q3
Income Q4
Income Q5
-2 -1 0 1 -2 -1 0 1 -2 -1 0 1 -2 -1 0 1 -2 -1 0 1 -2 -1 0 1
China Japan Korea Italy UK US
China Japan Korea Italy UK USCountry average 4.411 4.172 4.531 4.252 1.902 3.514
Note: Point estimates and 95% confidence intervals from a linear regression model for frequency of use of face masks on incomequintile, age group, gender and geographical controls. Figure based on regression results in Table A8. 18-25 and Income Q1 arebaseline categories for age and income quintile groups, respectively. Country averages of the outcome variable used for the regressionare reported below the figure.
Figure 4: Age and income gradients on frequency of use face mask
3.4 Support for the government and recommendations
The last question we turn to is the support of the population for the approach taken bytheir governments (Figure 5). Older individuals tend to be more supportive. Though wedo not observe such gradients in China—possibly because of ceiling effects, and Japan—where support for the government is the lowest. The pattern across income is again lessclear.
(a) Baseline specification
18-25
26-45
46-65
Above 65
Income Q1
Income Q2
Income Q3
Income Q4
Income Q5
-.5 0 .5 1 -.5 0 .5 1 -.5 0 .5 1 -.5 0 .5 1 -.5 0 .5 1 -.5 0 .5 1
China Japan Korea Italy UK US
204C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
(b) Specification with additional controls
18-25
26-45
46-65
Above 65
Income Q1
Income Q2
Income Q3
Income Q4
Income Q5
-.5 0 .5 1 -.5 0 .5 1 -.5 0 .5 1 -.5 0 .5 1 -.5 0 .5 1 -.5 0 .5 1
China Japan Korea Italy UK US
China Japan Korea Italy UK USCountry average 4.517 2.902 3.686 3.727 3.887 3.757
Note: Point estimates and 95% confidence intervals from a linear regression model of government support on income quintile, agegroup, gender, geographical controls, and additional controls. Additional controls include indicators for having lost the job at leasttemporarily and having lost household income, as well as count variables of the negative non-financial and positive non-financialeffects). 18-25 and Income Q1 are baseline categories for age and income quintile groups, respectively. Country averages of theoutcome variable used for the regression are reported below the figure.
Figure 5: Age and income gradients on support of the government’s handling of thepandemic
We explore to what extent age gradients can be explained by disproportionate effectsof the pandemic on different groups (using the variables presented above). When we con-trol for these variables (see bottom panel of Figure 5), we do not see substantial changesin these age and income gradients, suggesting that support is not directly driven by theeconomic, social or psychological impact of the outbreak on individuals.
4 Discussion
The epidemic and measures taken in response to it across the world appear to have af-fected different groups of the population in different ways. As a result, some subgroupsof the population are economically and psychologically more vulnerable than other sub-groups. Understanding the heterogeneous nature of the impact of Covid-19 impacts is anecessary step toward improving the current set of policy tools, i.e., to encourage com-pliance with measures that align with societal goals of containing the pandemic whileminimizing economic and social damage.
In the six countries we surveyed, we find consistent evidence that younger peoplehave been more negatively affected—both economically and psychologically—and thatthey appear to be less supportive of their governments’ approaches. On the other hand,we find a less clear pattern across income groups. Despite their lower ability to workfrom home, lower income people have not necessarily experienced the strongest negativeincome consequences, at least not in all countries. Some countries took early measures
205C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
to shield the low income groups from the economic consequences of the crisis. However,our findings on income groups are not definitive because the extent of economic impactswas not fully revealed by the time of our survey.
This evidence that younger people are more affected by the pandemic and supportless their government response strengthens the case for more differentiated policies thatshield the young from the negative consequences of the epidemic and necessary mea-sures. A number of recent papers propose policies that target lockdown policies to theolder part of the population (see, for the effects of age-specific policies, Acemoglu et al.,2020; Brotherhood et al., 2020; Favero et al., 2020). The advantage would be that such tar-geted policies would allow for economic recovery, while shielding those with the highesthealth risks. However, the consequences of shutting down interactions between the oldand the young are not yet well understood. People from different age groups rely on eachother for many reasons, and breaking such inter-generational bonds and arrangementsmay have negative consequences, which are difficult to assess and will require more em-pirical work. It is also imperative to find ways to match young people’s incentives andburdens of complying to public policies.
While our focus is on age and income differences, our data present important sys-tematic patterns indicating that women are disproportionately affected by this economiccrisis.5 For most of the surveyed countries women are less likely to have started tele-working, more likely to be socially isolated because of the pandemic, and more likely toreport suffering psychological consequences of the pandemic. These findings echo evi-dence presented by Alon et al. (2020) who show that women are concentrated in sectorsdisproportionately affected by the crisis. The data from Belot et al. (2020) do not includesome of the questions that are key for understanding the sources of such gender gaps (e.gtask allocation within the household), but a cross-country perspective can prove helpfulin directing ongoing investigations on the root causes of such gender gaps.
References
Acemoglu, Daron, Victor Chernozhukov, Ivan Werning, and Michael D Whinston, “Amulti-risk sir model with optimally targeted lockdown,” Technical Report, NationalBureau of Economic Research 2020.
Adams, Abigail, Teodora Boneva, Christopher Rauh, and Marta Golin, “Inequality inthe Impact of the Coronavirus Shock: New Survey Evidence for the UK,” 2020, (2023).
Alon, Titan M, Matthias Doepke, Jane Olmstead-Rumsey, and Michele Tertilt, “TheImpact of COVID-19 on Gender Equality,” Working Paper 26947, National Bureau ofEconomic Research April 2020.
Belot, Michele, Syngjoo Choi, Julian C Jamison, Nicholas W Papageorge, Egon Tripodi,and Eline van den Broek-Altenburg, “Six-Country Survey on Covid-19,” 2020.
5See also Etheridge et al. (2020) for evidence from the UK.
206C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Brotherhood, Luiz, Philipp Kircher, Cezar Santos, and Michele Tertilt, “An economicmodel of the Covid-19 epidemic: The importance of testing and age-specific policies,”2020.
Chetty, Raj, John N Friedman, Nathaniel Hendren, and Michael Stepner, “How DidCOVID-19 and Stabilization Policies Affect Spending and Employment? A New Real-Time Economic Tracker Based on Private Sector Data,” 2020.
Chu, Derek K, Elie A Akl, Stephanie Duda, Karla Solo, Sally Yaacoub, Holger JSchunemann, Amena El-harakeh, Antonio Bognanni, Tamara Lotfi, Mark Loeb et al.,“Physical distancing, face masks, and eye protection to prevent person-to-person trans-mission of SARS-CoV-2 and COVID-19: a systematic review and meta-analysis,” TheLancet, 2020.
Etheridge, Ben, Lisa Spantig et al., “The gender gap in mental well-being during theCovid-19 outbreak: evidence from the UK,” Technical Report, Institute for Social andEconomic Research 2020.
Fairlie, Robert W, Kenneth Couch, and Huanan Xu, “The Impacts of COVID-19 on Mi-nority Unemployment: First Evidence from April 2020 CPS Microdata,” Working Paper27246, National Bureau of Economic Research May 2020.
Favero, Carlo A, Andrea Ichino, and Aldo Rustichini, “Restarting the economy whilesaving lives under Covid-19,” 2020.
Hale, Thomas, Anna Petherick, Toby Phillips, and Samuel Webster, “Variation in gov-ernment responses to COVID-19,” Blavatnik School of Government Working Paper, 2020.
Howard, Jeremy, Austin Huang, Zhiyuan Li, Zeynep Tufekci, Vladimir Zdimal,Helene-Mari van der Westhuizen, Arne von Delft, Amy Price, Lex Fridman, Lei-HanTang et al., “Face masks against COVID-19: an evidence review,” 2020.
Montenovo, Laura, Xuan Jiang, Felipe Lozano Rojas, Ian M Schmutte, Kosali I Simon,Bruce A Weinberg, and Coady Wing, “Determinants of Disparities in Covid-19 JobLosses,” Working Paper 27132, National Bureau of Economic Research May 2020.
Papageorge, Nick, Matthew V. Zahn, Michele Belot, Eline van den Broek-Altenberg,Syngjoo Choi, Julian C. Jamison, and Egon Tripodi, “Socio-Demographic Factors As-sociated with Self-Protecting Behavior during the Covid-19 Pandemic,” 2020.
von Gaudecker, Hans-Martin, Radost Holler, Lena Janys, Bettina Siflinger, and Chris-tian Zimpelmann, “Labour supply in the early stages of the CoViD-19 Pandemic: Em-pirical Evidence on hours, home office, and expectations,” 2020.
207C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Appendix for Online Publication
A Additional Tables
Table A1: Linear probability model for having experienced household income loss
China Japan Korea Italy UK US 6 countries
Female -0.039 0.009 0.033 0.084∗∗∗ -0.018 0.035 0.012(0.031) (0.031) (0.033) (0.030) (0.030) (0.032) (0.013)
Age group (baseline: 18 to 25)
26 to 45 -0.116∗∗ -0.095∗ 0.021 0.069 0.014 0.084 -0.011(0.046) (0.055) (0.052) (0.051) (0.052) (0.053) (0.021)
46 to 65 -0.202∗∗∗ -0.096∗ 0.058 0.059 -0.124∗∗ -0.083 -0.073∗∗∗
(0.051) (0.055) (0.054) (0.052) (0.052) (0.052) (0.021)Above 66 -0.473∗∗∗ -0.271∗∗∗ -0.014 -0.276∗∗∗ -0.354∗∗∗ -0.259∗∗∗ -0.283∗∗∗
(0.063) (0.059) (0.063) (0.058) (0.060) (0.054) (0.024)
Income quintile (baseline: First quintile)
Second quintile -0.068 -0.027 -0.031 0.028 0.093∗ 0.014 0.003(0.049) (0.048) (0.053) (0.051) (0.050) (0.050) (0.021)
Third quintile -0.031 -0.050 -0.021 -0.030 0.167∗∗∗ 0.025 0.011(0.050) (0.048) (0.051) (0.049) (0.049) (0.049) (0.020)
Fourth quintile -0.093∗ -0.020 -0.064 -0.073 0.010 0.057 -0.024(0.051) (0.050) (0.051) (0.048) (0.048) (0.048) (0.020)
Fifth quintile -0.027 -0.069 -0.164∗∗∗ -0.132∗∗ 0.026 0.032 -0.038∗
(0.055) (0.053) (0.055) (0.054) (0.049) (0.052) (0.021)
Current living area (baseline: Urban)
Semi-urban 0.110∗∗∗ -0.108∗∗∗ -0.048 -0.012 -0.052 -0.027 -0.032∗∗
(0.039) (0.038) (0.048) (0.035) (0.035) (0.034) (0.015)Country-side 0.205∗∗∗ -0.094∗∗ -0.141∗ -0.123∗∗∗ 0.076 -0.132∗∗ -0.040∗
(0.061) (0.043) (0.083) (0.043) (0.050) (0.053) (0.021)
Constant 0.609∗∗∗ 0.654∗∗∗ 0.494∗∗∗ 0.552∗∗∗ 0.420∗∗∗ 0.505∗∗∗ 0.482∗∗∗
(0.083) (0.076) (0.078) (0.119) (0.078) (0.066) (0.069)
Regional fixed effects Y Y Y Y Y Y YCountry fixed effects N N N N N N Y
Observations 999 1013 964 1042 1016 1055 6089adj. R2 0.101 0.027 0.007 0.091 0.097 0.076 0.071
∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01Notes: Standard errors in parentheses. All specifications include regional fixed effects for the place of residence of the respondent(relevant administrative level is the province in China and South Korea, the region in Japan, Italy and the United Kingdom, and thestate in the United States).
208C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A2: Linear probability model for having experienced drop in spending
China Japan Korea Italy UK US 6 countries
Female 0.034 0.041 0.028 0.006 -0.097∗∗∗ -0.061∗ -0.016(0.032) (0.030) (0.033) (0.031) (0.030) (0.032) (0.013)
Age group (baseline: 18 to 25)
26 to 45 -0.047 -0.091∗ -0.032 0.096∗ -0.002 0.044 -0.011(0.049) (0.053) (0.052) (0.053) (0.052) (0.053) (0.021)
46 to 65 -0.120∗∗ -0.184∗∗∗ 0.019 0.067 -0.030 -0.061 -0.055∗∗∗
(0.054) (0.052) (0.053) (0.054) (0.053) (0.052) (0.021)Above 66 -0.289∗∗∗ -0.203∗∗∗ -0.034 -0.024 -0.129∗∗ -0.149∗∗∗ -0.138∗∗∗
(0.067) (0.056) (0.063) (0.061) (0.060) (0.054) (0.024)
Income quintile (baseline: First quintile)
Second quintile 0.080 0.010 -0.029 0.032 0.118∗∗ 0.119∗∗ 0.054∗∗∗
(0.052) (0.046) (0.053) (0.053) (0.051) (0.050) (0.021)Third quintile 0.032 0.022 -0.090∗ -0.018 0.248∗∗∗ 0.129∗∗∗ 0.058∗∗∗
(0.053) (0.046) (0.051) (0.051) (0.049) (0.049) (0.020)Fourth quintile -0.006 0.025 -0.003 0.038 0.152∗∗∗ 0.168∗∗∗ 0.074∗∗∗
(0.053) (0.048) (0.051) (0.051) (0.048) (0.048) (0.020)Fifth quintile -0.090 0.009 -0.053 -0.032 0.185∗∗∗ 0.232∗∗∗ 0.058∗∗∗
(0.058) (0.051) (0.055) (0.056) (0.049) (0.052) (0.021)
Current living area (baseline: Urban)
Semi-urban 0.016 -0.001 -0.069 -0.025 -0.046 0.031 -0.020(0.041) (0.036) (0.048) (0.037) (0.035) (0.034) (0.015)
Country-side -0.014 -0.037 -0.104 -0.086∗ 0.015 -0.039 -0.044∗∗
(0.064) (0.041) (0.083) (0.045) (0.050) (0.053) (0.021)
Constant 0.519∗∗∗ 0.380∗∗∗ 0.487∗∗∗ 0.427∗∗∗ 0.325∗∗∗ 0.407∗∗∗ 0.455∗∗∗
(0.087) (0.073) (0.078) (0.124) (0.078) (0.065) (0.069)
Regional fixed effects Y Y Y Y Y Y YCountry fixed effects N N N N N N Y
Observations 999 1013 964 1042 1016 1055 6089adj. R2 0.027 0.012 0.018 0.017 0.068 0.062 0.055
∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01Notes: Standard errors in parentheses. All specifications include regional fixed effects for the place of residence of the respondent(relevant administrative level is the province in China and South Korea, the region in Japan, Italy and the United Kingdom, andthe state in the United States).
209C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A3: Ordinary least squares for having started teleworking
China Japan Korea Italy UK US 6 countries
Female -0.007 -0.045∗ -0.046∗∗ -0.027 -0.102∗∗∗ -0.080∗∗∗ -0.057∗∗∗
(0.029) (0.023) (0.023) (0.026) (0.023) (0.024) (0.010)
Age group (baseline: 18 to 25)
26 to 45 -0.111∗∗ -0.050 -0.055 0.025 0.015 0.157∗∗∗ -0.000(0.044) (0.041) (0.037) (0.045) (0.040) (0.040) (0.017)
46 to 65 -0.252∗∗∗ -0.108∗∗∗ -0.120∗∗∗ -0.022 -0.089∗∗ 0.001 -0.092∗∗∗
(0.048) (0.041) (0.037) (0.046) (0.041) (0.039) (0.017)Above 66 -0.574∗∗∗ -0.242∗∗∗ -0.136∗∗∗ -0.235∗∗∗ -0.193∗∗∗ -0.141∗∗∗ -0.242∗∗∗
(0.060) (0.044) (0.044) (0.051) (0.046) (0.041) (0.019)
Income quintile (baseline: First quintile)
Second quintile 0.211∗∗∗ 0.084∗∗ 0.034 0.066 0.025 0.080∗∗ 0.094∗∗∗
(0.047) (0.036) (0.037) (0.045) (0.039) (0.038) (0.016)Third quintile 0.075 0.156∗∗∗ 0.145∗∗∗ 0.161∗∗∗ 0.095∗∗ 0.153∗∗∗ 0.145∗∗∗
(0.047) (0.036) (0.035) (0.043) (0.038) (0.037) (0.016)Fourth quintile 0.124∗∗∗ 0.214∗∗∗ 0.130∗∗∗ 0.229∗∗∗ 0.152∗∗∗ 0.259∗∗∗ 0.201∗∗∗
(0.048) (0.038) (0.036) (0.042) (0.037) (0.036) (0.016)Fifth quintile 0.171∗∗∗ 0.327∗∗∗ 0.162∗∗∗ 0.351∗∗∗ 0.219∗∗∗ 0.304∗∗∗ 0.278∗∗∗
(0.052) (0.040) (0.038) (0.047) (0.038) (0.040) (0.017)
Current living area (baseline: Urban)
Semi-urban -0.118∗∗∗ -0.010 -0.029 -0.102∗∗∗ -0.041 -0.041 -0.073∗∗∗
(0.037) (0.028) (0.034) (0.031) (0.027) (0.026) (0.012)Country-side -0.249∗∗∗ -0.029 -0.038 -0.127∗∗∗ -0.059 -0.128∗∗∗ -0.110∗∗∗
(0.058) (0.032) (0.058) (0.038) (0.039) (0.040) (0.017)
Constant 0.674∗∗∗ 0.143∗∗ 0.154∗∗∗ 0.222∗∗ 0.163∗∗∗ 0.107∗∗ 0.508∗∗∗
(0.078) (0.057) (0.054) (0.104) (0.060) (0.050) (0.055)
Regional fixed effects Y Y Y Y Y Y YCountry fixed effects N N N N N N Y
Observations 999 1013 964 1042 1016 1055 6089adj. R2 0.229 0.149 0.036 0.126 0.110 0.259 0.213
∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01Notes: Standard errors in parentheses. All specifications include regional fixed effects for the place of residence of the respondent(relevant administrative level is the province in China and South Korea, the region in Japan, Italy and the United Kingdom, and thestate in the United States).
210C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A4: Ordinary least squares for index of dissatisfaction with social distance
China Japan Korea Italy UK US 6 countries
Female -0.266∗∗∗ 0.366∗∗∗ 0.129 0.024 0.083 0.063 0.067∗
(0.077) (0.107) (0.109) (0.084) (0.098) (0.100) (0.039)
Age group (baseline: 18 to 25)
26 to 45 -0.170 -0.124 -0.202 0.181 -0.268 0.088 -0.102(0.116) (0.190) (0.173) (0.144) (0.169) (0.165) (0.065)
46 to 65 -0.256∗∗ -0.324∗ -0.226 -0.002 -0.226 -0.098 -0.184∗∗∗
(0.128) (0.188) (0.177) (0.147) (0.171) (0.163) (0.065)Above 66 -0.342∗∗ -0.947∗∗∗ -0.731∗∗∗ -0.238 -0.523∗∗∗ -0.272 -0.519∗∗∗
(0.158) (0.202) (0.208) (0.163) (0.195) (0.171) (0.073)
Income quintile (baseline: First quintile)
Second quintile 0.325∗∗∗ 0.253 0.462∗∗∗ -0.007 0.425∗∗∗ 0.331∗∗ 0.286∗∗∗
(0.124) (0.165) (0.175) (0.144) (0.165) (0.156) (0.063)Third quintile 0.359∗∗∗ 0.258 0.668∗∗∗ 0.171 0.430∗∗∗ 0.313∗∗ 0.337∗∗∗
(0.126) (0.164) (0.168) (0.138) (0.160) (0.153) (0.062)Fourth quintile 0.337∗∗∗ 0.412∗∗ 0.895∗∗∗ 0.367∗∗∗ 0.576∗∗∗ 0.431∗∗∗ 0.500∗∗∗
(0.127) (0.172) (0.169) (0.136) (0.156) (0.151) (0.061)Fifth quintile 0.388∗∗∗ 0.558∗∗∗ 0.803∗∗∗ 0.444∗∗∗ 0.550∗∗∗ 0.386∗∗ 0.511∗∗∗
(0.137) (0.182) (0.181) (0.152) (0.159) (0.165) (0.065)
Current living area (baseline: Urban)
Semi-urban 0.203∗∗ 0.031 -0.334∗∗ 0.124 -0.035 -0.006 0.015(0.098) (0.130) (0.159) (0.099) (0.114) (0.107) (0.047)
Country-side -0.053 0.032 -0.672∗∗ -0.166 0.010 0.144 -0.033(0.153) (0.147) (0.274) (0.122) (0.164) (0.168) (0.064)
Constant -0.478∗∗ -0.160 -0.381 0.033 -0.265 -0.132 -0.646∗∗∗
(0.207) (0.261) (0.257) (0.334) (0.255) (0.206) (0.212)
Regional fixed effects Y Y Y Y Y Y YCountry fixed effects N N N N N N Y
Observations 999 1013 964 1042 1016 1055 6089adj. R2 0.075 0.049 0.052 0.040 0.042 0.016 0.037
∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01Notes: Standard errors in parentheses. All specifications include regional fixed effects for the place of residence of the respondent(relevant administrative level is the province in China and South Korea, the region in Japan, Italy and the United Kingdom, andthe state in the United States). The outcome is index that is constructed as the first principal component fitting the variables thatcapture how bothered respondents are for not being able to meet other people in their free time, do leisure activities outside ofhome, shop non-essentials.
211C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A5: Ordinary least square for negative non-financial effects
China Japan Korea Italy UK US 6 countries
Female 0.024 0.366∗∗∗ 0.241∗∗∗ 0.281∗∗∗ 0.387∗∗∗ 0.404∗∗∗ 0.277∗∗∗
(0.068) (0.068) (0.071) (0.074) (0.085) (0.085) (0.030)
Age group (baseline: 18 to 25)
26 to 45 -0.302∗∗∗ -0.317∗∗∗ -0.159 -0.379∗∗∗ -0.318∗∗ -0.219 -0.284∗∗∗
(0.103) (0.121) (0.113) (0.127) (0.145) (0.141) (0.051)46 to 65 -0.573∗∗∗ -0.600∗∗∗ -0.367∗∗∗ -0.577∗∗∗ -0.589∗∗∗ -0.792∗∗∗ -0.568∗∗∗
(0.113) (0.119) (0.115) (0.130) (0.147) (0.139) (0.051)Above 66 -0.714∗∗∗ -0.813∗∗∗ -0.457∗∗∗ -0.891∗∗∗ -1.207∗∗∗ -1.134∗∗∗ -0.885∗∗∗
(0.140) (0.128) (0.136) (0.144) (0.168) (0.146) (0.058)
Income quintile (baseline: First quintile)
Second quintile -0.012 0.131 -0.048 0.142 -0.131 0.122 0.023(0.110) (0.105) (0.114) (0.127) (0.141) (0.133) (0.050)
Third quintile 0.154 0.040 0.084 -0.028 -0.157 0.085 0.002(0.111) (0.104) (0.109) (0.122) (0.138) (0.131) (0.048)
Fourth quintile -0.015 0.192∗ -0.058 -0.002 -0.331∗∗ 0.135 -0.027(0.112) (0.109) (0.110) (0.120) (0.134) (0.129) (0.048)
Fifth quintile 0.107 0.203∗ -0.029 -0.061 -0.253∗ -0.049 -0.026(0.121) (0.116) (0.118) (0.134) (0.137) (0.141) (0.051)
Current living area (baseline: Urban)
Semi-urban 0.341∗∗∗ -0.053 -0.044 0.026 0.099 0.081 0.066∗
(0.087) (0.082) (0.104) (0.088) (0.098) (0.091) (0.037)Country-side -0.011 -0.046 -0.330∗ -0.372∗∗∗ -0.057 -0.060 -0.138∗∗∗
(0.135) (0.094) (0.178) (0.108) (0.141) (0.143) (0.050)
Constant 1.622∗∗∗ 1.446∗∗∗ 1.644∗∗∗ 1.811∗∗∗ 1.903∗∗∗ 1.939∗∗∗ 1.624∗∗∗
(0.183) (0.166) (0.167) (0.295) (0.219) (0.176) (0.166)
Regional fixed effects Y Y Y Y Y Y YCountry fixed effects N N N N N N Y
Observations 999 1013 964 1042 1016 1055 6089adj. R2 0.086 0.083 0.030 0.087 0.088 0.105 0.092
∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01Notes: Standard errors in parentheses. All specifications include regional fixed effects for the place of residence of the respondent(relevant administrative level is the province in China and South Korea, the region in Japan, Italy and the United Kingdom, andthe state in the United States). The outcome is a variable that counts how many of the following non-financial negative effectsrespondents report to be experiencing due to the pandemic. This includes (i) boredom, (ii) loneliness, (iii) trouble sleeping, (iv)general anxiety and stress, (v) increased conflicts with friends/relatives/neighbors.
212C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A6: Ordinary least squares for index of belief of policy effectiveness (higher valuesdenote belief in higher effectiveness)
China Japan Korea Italy UK US 6 countries
Female 0.209∗ 0.442∗∗∗ 0.435∗∗∗ 0.247∗ 0.561∗∗∗ 0.509∗∗∗ 0.387∗∗∗
(0.123) (0.145) (0.110) (0.132) (0.125) (0.144) (0.052)
Age group (baseline: 18 to 25)
26 to 45 0.137 0.021 0.212 0.449∗∗ 0.504∗∗ 0.236 0.259∗∗∗
(0.186) (0.257) (0.173) (0.226) (0.216) (0.240) (0.087)46 to 65 0.106 -0.000 0.355∗∗ 0.592∗∗ 0.731∗∗∗ 0.428∗ 0.333∗∗∗
(0.204) (0.255) (0.178) (0.230) (0.218) (0.236) (0.088)Above 66 0.245 0.250 0.556∗∗∗ 0.781∗∗∗ 0.792∗∗∗ 0.597∗∗ 0.529∗∗∗
(0.253) (0.274) (0.209) (0.256) (0.248) (0.247) (0.099)
Income quintile (baseline: First quintile)
Second quintile -0.384∗ -0.363 0.570∗∗∗ 0.345 0.174 -0.177 0.004(0.198) (0.225) (0.175) (0.226) (0.209) (0.226) (0.086)
Third quintile -0.250 0.314 0.583∗∗∗ 0.332 0.266 0.251 0.240∗∗∗
(0.201) (0.223) (0.168) (0.216) (0.204) (0.222) (0.083)Fourth quintile 0.238 -0.075 0.526∗∗∗ 0.209 0.062 0.404∗ 0.216∗∗∗
(0.203) (0.233) (0.169) (0.214) (0.199) (0.218) (0.083)Fifth quintile 0.010 -0.263 0.596∗∗∗ 0.090 0.109 0.533∗∗ 0.182∗∗
(0.220) (0.247) (0.181) (0.238) (0.203) (0.238) (0.088)
Current living area (baseline: Urban)
Semi-urban 0.085 0.114 -0.138 0.081 -0.053 -0.210 -0.019(0.157) (0.176) (0.159) (0.156) (0.145) (0.155) (0.063)
Country-side 0.233 0.011 0.017 0.066 0.049 -0.258 0.014(0.245) (0.200) (0.275) (0.191) (0.209) (0.243) (0.087)
Constant 0.283 -0.832∗∗ -1.743∗∗∗ -0.295 -1.425∗∗∗ -0.806∗∗∗ -0.101(0.331) (0.354) (0.258) (0.524) (0.325) (0.298) (0.287)
Regional fixed effects Y Y Y Y Y Y YCountry fixed effects N N N N N N Y
Observations 999 1013 964 1042 1016 1055 6089adj. R2 0.005 0.013 0.030 -0.002 0.015 0.021 0.053
∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01Notes: Standard errors in parentheses. All specifications include regional fixed effects for the place of residence of the respon-dent (relevant administrative level is the province in China and South Korea, the region in Japan, Italy and the United Kingdom,and the state in the United States).
213C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A7: Ordinary least squares for positive non-financial effects
China Japan Korea Italy UK US 6 countries
Female -0.003 -0.018 0.101∗ -0.010 0.075 0.082 0.025(0.062) (0.051) (0.059) (0.063) (0.069) (0.069) (0.025)
Age group (baseline: 18 to 25)
26 to 45 -0.066 -0.140 -0.188∗∗ 0.005 -0.227∗ 0.064 -0.100∗∗
(0.094) (0.090) (0.093) (0.108) (0.118) (0.114) (0.042)46 to 65 -0.430∗∗∗ -0.153∗ -0.273∗∗∗ 0.093 0.018 -0.099 -0.127∗∗∗
(0.103) (0.089) (0.095) (0.110) (0.120) (0.112) (0.042)Above 66 -0.546∗∗∗ -0.092 -0.371∗∗∗ 0.113 0.103 -0.244∗∗ -0.162∗∗∗
(0.128) (0.096) (0.111) (0.122) (0.136) (0.117) (0.047)
Income quintile (baseline: First quintile)
Second quintile 0.024 0.154∗∗ 0.210∗∗ 0.132 -0.136 0.152 0.101∗∗
(0.100) (0.079) (0.093) (0.108) (0.115) (0.108) (0.041)Third quintile 0.076 0.188∗∗ 0.134 0.229∗∗ -0.111 0.248∗∗ 0.144∗∗∗
(0.102) (0.078) (0.090) (0.104) (0.112) (0.105) (0.040)Fourth quintile -0.055 0.190∗∗ 0.191∗∗ 0.147 0.030 0.141 0.134∗∗∗
(0.103) (0.082) (0.090) (0.102) (0.109) (0.104) (0.040)Fifth quintile 0.101 0.218∗∗ 0.023 0.259∗∗ 0.015 0.095 0.158∗∗∗
(0.111) (0.086) (0.097) (0.114) (0.111) (0.113) (0.042)
Current living area (baseline: Urban)
Semi-urban 0.296∗∗∗ -0.171∗∗∗ 0.014 0.149∗∗ 0.078 0.069 0.055∗
(0.080) (0.061) (0.085) (0.075) (0.079) (0.073) (0.030)Country-side 0.163 -0.163∗∗ 0.242∗ -0.083 0.051 0.016 0.016
(0.124) (0.070) (0.147) (0.092) (0.115) (0.115) (0.042)Constant 1.740∗∗∗ 0.518∗∗∗ 0.827∗∗∗ 1.541∗∗∗ 1.329∗∗∗ 1.220∗∗∗ 1.528∗∗∗
(0.167) (0.124) (0.137) (0.251) (0.178) (0.142) (0.137)
Regional fixed effects Y Y Y Y Y Y YCountry fixed effects N N N N N N Y
Observations 999 1013 964 1042 1016 1055 6089adj. R2 0.060 0.025 0.019 0.016 0.006 0.018 0.115
∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01Notes: Standard errors in parentheses. All specifications include regional fixed effects for the place of residence of the respon-dent (relevant administrative level is the province in China and South Korea, the region in Japan, Italy and the United Kingdom,and the state in the United States). The outcome is a variable that counts how many of the following non-financial positiveeffects respondents report to be experiencing due to the pandemic. This includes (i) enjoying more free time, (ii) enjoying timewith family, (iii) reduction of air pollution, (iv) reduction of noise pollution.
214C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Table A8: Ordinary least squares for frequency of wearing masks “now” (i.e. around timeof data collection)
China Japan Korea Italy UK US 6 countries
Female 0.133∗∗ 0.663∗∗∗ 0.212∗∗∗ 0.110 -0.174∗∗ 0.151 0.177∗∗∗
(0.064) (0.076) (0.062) (0.078) (0.086) (0.099) (0.032)
Age group (baseline: 18 to 25)
26 to 45 -0.070 0.070 -0.058 0.474∗∗∗ -0.058 0.140 0.087(0.097) (0.134) (0.098) (0.135) (0.148) (0.163) (0.054)
46 to 65 -0.089 0.029 0.039 0.418∗∗∗ -0.627∗∗∗ 0.067 -0.049(0.107) (0.133) (0.100) (0.137) (0.149) (0.161) (0.054)
Above 66 -0.008 0.210 0.021 0.565∗∗∗ -0.946∗∗∗ 0.237 0.004(0.132) (0.143) (0.118) (0.152) (0.171) (0.169) (0.061)
Income quintile (baseline: First quintile)
Second quintile 0.081 -0.059 0.411∗∗∗ 0.074 0.125 0.199 0.117∗∗
(0.104) (0.117) (0.099) (0.135) (0.144) (0.154) (0.052)Third quintile 0.133 -0.015 0.389∗∗∗ 0.042 0.246∗ 0.500∗∗∗ 0.199∗∗∗
(0.105) (0.116) (0.095) (0.129) (0.140) (0.151) (0.051)Fourth quintile 0.082 0.186 0.328∗∗∗ -0.000 -0.034 0.463∗∗∗ 0.157∗∗∗
(0.106) (0.122) (0.096) (0.127) (0.136) (0.149) (0.051)Fifth quintile 0.214∗ 0.116 0.448∗∗∗ 0.026 -0.042 0.677∗∗∗ 0.249∗∗∗
(0.115) (0.129) (0.102) (0.142) (0.139) (0.162) (0.054)
Current living area (baseline: Urban)
Semi-urban 0.123 -0.116 0.048 0.012 -0.280∗∗∗ -0.422∗∗∗ -0.150∗∗∗
(0.082) (0.092) (0.090) (0.093) (0.099) (0.105) (0.039)Country-side -0.090 -0.050 -0.217 0.025 -0.122 -0.641∗∗∗ -0.178∗∗∗
(0.128) (0.104) (0.155) (0.114) (0.143) (0.166) (0.053)
Constant 4.330∗∗∗ 3.728∗∗∗ 4.040∗∗∗ 3.843∗∗∗ 2.471∗∗∗ 3.267∗∗∗ 4.280∗∗∗
(0.173) (0.185) (0.145) (0.312) (0.223) (0.203) (0.176)
Regional fixed effects Y Y Y Y Y Y YCountry fixed effects N N N N N N Y
Observations 999 1013 964 1042 1016 1055 6089adj. R2 0.011 0.075 0.045 0.002 0.116 0.056 0.363
∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01Notes: Standard errors in parentheses. All specifications include regional fixed effects for the place of residence of the respon-dent (relevant administrative level is the province in China and South Korea, the region in Japan, Italy and the United Kingdom,and the state in the United States).
215C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
0
20
40
60
80
100
0
20
40
60
80
100
0
1000
2000
3000
4000
0
1000
2000
3000
4000
01-01 02-01 03-01 04-01 05-01 01-01 02-01 03-01 04-01 05-01 01-01 02-01 03-01 04-01 05-01
China Japan Korea
Italy UK US
Cases Stringency
Stri
ngen
cy in
dex
Cas
es p
er m
illio
n in
habi
tant
s
Date
Note: The gray bar represents the third week of April 2020, in which the survey was conducted. Source: Hale et al. (2020).
Figure A1: Time series of the number of confirmed cases and stringency index of govern-ment responses
216C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
0
20
40
60
80
100
0
20
40
60
80
100
0
100
200
300
400
500
0
100
200
300
400
500
01-01 02-01 03-01 04-01 05-01 01-01 02-01 03-01 04-01 05-01 01-01 02-01 03-01 04-01 05-01
China Japan Korea
Italy UK US
Deaths Stringency
Stri
ngen
cy in
dex
Dea
ths
per
mill
ion
inha
bita
nts
Date
Note: The gray bar represents the third week of April 2020, in which the survey was conducted. Data on deaths from China is notavailable. Source: Hale et al. (2020).
Figure A2: Time series of the number of Covid-19 related deaths and stringency index ofgovernment responses
217C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
196
-217
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
Covid Economics Issue 38, 16 July 2020
Copyright: Raymundo M. Campos-Vazquez and Gerardo Esquivel
Consumption and geographic mobility in pandemic times: Evidence from Mexico1
Raymundo M. Campos-Vazquez2 and Gerardo Esquivel3
Date submitted: 11 July 2020; Date accepted: 14 July 2020
We analyze the universe of point-of-sale (POS) transactions before and during the COVID-19 lockdown in Mexico. We find three key results. First, consumption in Mexico fell by 23 percent in the April-June quarter of 2020. Second, reductions in consumption were highly heterogeneous across sectors and states, with states and activities related to tourism the most affected. Third, using variation over time and states, we estimate the elasticity of POS expenditures with respect to geographic mobility (measured using cellphone location data) to be slightly less than 1. This estimate suggests that spending in developing countries may be more responsive to mobility than in developed countries, and that mobility indicators could be used as a real-time proxy for consumption in some economies.
1 We are grateful for superb research assistance from Raquel Yunoen Badillo and Kavik Rocha. The content of this study does not necessarily reflect the official opinion of the Banco de Mexico or its Board. Responsibility for the information, errors, omissions, and views expressed here lies entirely with the authors.
2 Professor, El Colegio de Mexico (on academic leave at the Banco de Mexico).3 Banco de México.
218C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
1
1. INTRODUCTION
The economic consequences of COVID-19 are significant. Lockdown and health
measures have substantially decreased geographic mobility, causing a drop in economic
activity. The traditional economic indicators that measure these effects, like gross
domestic product (GDP) and the industrial production index (IPI) are published by
national statistical agencies with a lag: in Mexico, approximately two months after the
fact. Researchers and policy-makers across the world are trying to overcome this delay
by analyzing high-frequency data to quantify the magnitude of the shock and make
prescriptions to avoid a more severe economic contraction (see, for example, the weekly
economic index of Lewis, Mertens, and Stock 2020; the index of expenditures of Baker
et al. 2020; and the labor market index of Kahn, Lange, and Wiczer 2020). Given the
possibility of future waves of COVID-19, it is extremely important to measure the
relationship of mobility and economic activity. In this paper, we use aggregated daily
point-of-sale (POS) transaction data and cellphone location data in Mexico to quantify
the magnitude of the shock and to estimate the effect of mobility patterns on POS
expenditures.
It is now well known that a supply shock may cause a demand shock in the
economy, thus amplifying the initial economic impact (Guerrieri et al. 2020). Sectors
related to services, such as restaurants and tourism, are directly affected by a pandemic.
One could then expect that the total shock should be proportional to the income losses of
these sectors. However, income generated in other sectors may be affected as well,
depending on the value of current versus future consumption and the value of goods and
services not provided during the pandemic. If we have a high intertemporal elasticity of
substitution (e.g., people can modify their consumption patterns relatively easily to spend
more later rather than now) and a low intra-temporal elasticity of substitution (e.g., people
219C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
2
prefer to buy the same goods and services, and there are no good substitutes for their
consumption patterns), then a demand shock exacerbates the original shock, which can
present an even greater problem in the presence of uncertainty and incomplete markets.
It is thus important to estimate how total expenditure is changing over time and
which sectors are most affected, an estimate that requires high frequency data. POS
expenditure data may meet the requirements for such analysis. Indeed, there are recent
articles that make use of such information. In the United States, Baker et al. (2020) use
de-identified non-random data from a Fintech company at the transaction and individual
level. They find a spike in total spending when cases begin to increase (late February and
early March) but a subsequent decrease of close to 50 percent with respect to January and
early February. In Spain, Carvalho et al. (2020) use all POS transactions of customers of
a commercial bank and transactions of others using the POS terminals of that bank. As in
the U.S. study, they find a spike before the mid-March lockdown and then a sharp decline
in total expenditure: 60 percent with respect to the same period in 2019. In Denmark,
Andersen et al. (2020a) use data from the country’s largest retail bank. They find a
decrease in total spending of around 25 percent after lockdown starts. Similar results have
been found in other countries: the United Kingdom shows a decline of 46 percent from
April 2019 to April 2020 (Hacioglu, Känzig, and Surico 2020), France a decline of 60
percent (Bounie, Camara, and Galbraith 2020), Portugal a reduction of 55 percent in total
purchases in April (Carvalho, Peralta, and Pereira 2020), and China a decline of 42
percent (Chen, Qian, and Wen 2020).
POS data is useful for shedding light on causes and potential solutions for the
current crisis. Using U.S. data, Chetty et al. (2020) argue that the drop in POS
expenditures is driven mainly by rich households due to health concerns. Expenditures in
poor households generally returned to 2019 levels after their stimulus payments arrived.
220C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
3
Employment losses are greater in higher-income zip codes, especially in personal services
like restaurants and barber shops. They conclude that economic recovery goes hand in
hand with safety concerns.
Our paper makes important contributions to this literature. First, we show that the
response in developing countries may be different than in developed countries. Although
Mexico is an upper middle-income country, its financial sector is not as developed as in
other countries. According to the World Bank (2020), domestic bank credit to the private
sector accounts for only 27 percent of GDP, while in countries with similar consumption
patterns, like China, Denmark, France, Spain, and the United Kingdom, it is close to or
above 100 percent. Only in the United States is it less than that, and even there it is 52
percent of GDP. Also, the number of POS terminals in Mexico per 100,000 population is
the lowest among similar countries (approximately 1000 in Mexico versus 2000 in China
and 3000 in the other countries). Finally, internet penetration in Mexico (around 64
percent) is less than in the United States (76 percent) or similar European countries (all
above 80 percent). Although this may mean that POS data are not as comprehensive for
Mexico, our results indicate large negative effects of the pandemic, although not as large
as in those in other countries.
Second, the data we analyze for Mexico includes all POS transactions in the
country, in contrast to the data in previous studies, which is limited to selected banks or
companies. The comprehensive nature of our data allows us to benchmark the effect of
COVID-19 on POS expenditures to traditional measures like total consumption and GDP.
Third, although we follow previous literature in calculating expenditure losses with
respect to 2019, we also propose a simple model to calculate a counterfactual of what
expenditure would have looked like in the absence of the pandemic. Finally, we estimate
the elasticity of POS expenditures with respect to measures of geographic mobility using
221C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
4
variation over time within states in Mexico. This elasticity is important, as it could be
used in theoretical models and simulation exercises to calculate expenditure losses for
future waves of the pandemic. It is also an important consideration in the debate about
the impact of lockdown measures on the level of expenditures.
We use the universe of point-of-sale (POS) transactions from January 1, 2019 to
June 30, 2020, which is non-public data from the Banco de México (the Mexican central
bank), consisting of aggregated daily information on total expenditures and certain other
categories. This POS expenditure data provides important information about general
consumption patterns. In 2019, there were 157 million debit and credit cards in Mexico,
and the National Financial Inclusion Survey (INEGI, 2018) shows that more than two-
thirds of the Mexican population (68 percent) aged 18-70 have at least one such financial
product. In 2019, the average POS expenditure per transaction was $630 MXN
(approximately $31 USD). Approximately 10 million transactions take place through
POS terminals every day, 73 percent of which are with debit cards and the remaining 27
percent with credit cards. The average monthly total debit and credit card expenditure
was almost $187 billion MXN during 2019 (approximately $9.2 billion USD). Annual
total POS expenditure thus represents about 8 percent of GDP and 14 percent of
consumption.
We are able to provide the first direct estimates of the elasticity of POS
expenditures with respect to geographic mobility. Previous studies have provided only
indirect or implicit estimates for this elasticity. For example, using the results in Andersen
et al. (2020b), we can estimate an elasticity of 0.2 by exploiting the between-country
variation in spending and mobility for Sweden and Denmark: consumption declined 29
percent in Denmark and by 25 percent in Sweden (Figure 3 in that study). Using mobility
measures based on cellphone location data available from the Apple Corporation (2020)
222C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
5
for early April, we find that mobility decreased by only 12 percent in Sweden while it
declined by 32 percent in Denmark. The implicit elasticity of POS expenditures with
respect to mobility is thus around 0.2. In the current study, since we have daily data for
expenditures in Mexico at the subnational level, we are able to estimate the elasticity of
consumption with respect to mobility indicators by exploiting both the time and
geographic variation in the data.
We find three key results. First, the percent loss in POS expenditures with respect
to the estimate without the pandemic is 23 percent for April-June. This estimate is much
lower than that for other countries. The estimate for Spain and France (for the last two
weeks of March) is close to 50 percent (Bounie et al. 2020; Carvalho et al. 2020), for
Portugal it is 55 percent (Carvalho, Peralta, and Pereira 2020), and for Denmark it is 30
percent (Andersen et al. 2020a). Although estimates for the U.S. vary, our result is similar
to the live results from POS data in Chetty et al. (2020). In terms of GDP and
consumption, for the April-June quarter it implies a loss of 2.6 percent of quarterly GDP
and 3.9 percent of quarterly private consumption.
Second, losses vary significantly across sectors and regions. While some sectors
were severely hit, like tourism, food services, and transportation, others, like insurance
and telecommunications, were barely affected. This result is similar to that found in other
studies. Mexican states that are highly dependent on tourism (beach resorts and other
tourist destinations) are among the most affected.
Third, we estimate the elasticity of POS expenditures with respect to geographic
mobility in Mexico, as measured using cellphone location data from Apple (2020) and
Google (2020). Our estimates show that this elasticity is in most cases non-significantly
different from one (0.93 using Apple’s measure of mobility in one specification, and 0.91
for both Google’s and Apple’s measures of mobility in another). These estimates are
223C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
6
much larger and more precisely estimated than the estimate of 0.2 derived by comparing
the effect of mobility on spending in Sweden and Denmark, as described above. This
result suggests that POS expenditures in developing countries could be more responsive
to mobility patterns than in developed countries, an interesting possibility that calls for
further research. It may be possible, for example, that internet penetration and the strength
of e-commerce affect the magnitude of this elasticity. This result is also important because
it suggests that in economies like Mexico’s, mobility indicators, which can be observed
almost in real time, could serve as a good proxy for the behavior of expenditures.
2. DATA AND METHODS
The data includes all point-of-sale (POS) transactions in Mexican territory, which
is non-public information collected by the Banco de México under its mandate to assure
a well-functioning payment system. The data is aggregated by type of card (debit or
credit), at the state and national levels, and by type of expenditure, on a daily basis, from
January 2019 to June 2020. We observe only aggregate information; we do not observe
any individual transactions, any information about whether the credit or debit card is
foreign or Mexican, or whether the transaction took place on the internet or in a physical
location.
Most of the previous literature uses either a part of the universe of transactions or
a sample of households. Our use of the full universe of transactions allows us to calculate
total losses in the economy. However, one key challenge is how to construct a valid
counterfactual for comparison. In general, previous studies calculate the percent change
in 2020 with respect to 2019. This seems reasonable if the financial sector is stable.
However, because transactions in Mexico were already growing before the pandemic
arrived it seems more appropriate to construct a counterfactual scenario using data from
224C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
7
2019 and 2020. We propose a simple model that predicts the daily (t) outcome in 2020
(𝑃𝑂𝑆 ) based on both the 2019 outcome (𝑃𝑂𝑆 ) and pre-pandemic data observed
for 2020.1 We also include dummy variables related to paydays, Mondays, Fridays, and
for the month of December.
𝑃𝑂𝑆 𝛼 𝛽𝑃𝑂𝑆 𝑑𝑢𝑚𝑚𝑖𝑒𝑠 𝜀 (1)
The regression is estimated for all days from January 1, 2020 to February 18, 2020.
We select the final model minimizing the mean squared error for the prediction for
February 19 to March 11, 2020, that is, during the pre-lockdown period. Then, we make
a prediction for all the remaining days in 2020. All of the predictions are in constant pesos
(MXN) of July 2018. The percent effect of the pandemic can then be calculated as:
∆% 𝐸𝑓𝑓𝑒𝑐𝑡 (2)
The comparison with respect to 2019 replaces the predicted value 𝑃𝑂𝑆 with the value
in 2019, 𝑃𝑂𝑆 . As the daily expenditures are noisy, in some cases we smooth the lines
in the figures by a simple moving average for the previous two weeks. We show below
multiple estimates for total expenditures, for credit and debit cards, for type of
expenditure, and at the state level.
We also calculate the elasticity of total expenditures with respect to indicators of
geographic mobility, obtained from Google (2020) and Apple (2020). Google tracks
mobility using the location history of the Google accounts on people’s mobile devices;
we use this data to calculate the percent change compared with the median value for
baseline days in the five-week period January 3 to February 6, 2020. We focus on the
mobility trends for workplaces. Apple mobility is an index with a baseline set at January
1 We compare this model with other ARIMA models with the form 𝑃𝑂𝑆 𝛼 𝛽𝑃𝑂𝑆 𝜃𝐴𝑅𝐼𝑀𝐴 𝑑𝑢𝑚𝑚𝑖𝑒𝑠 𝜀 in terms of the root mean squared predicted error (RMSPE) for February 19 to March 11. This comparison is for total, credit, and debit expenditures, varying the introduction of dummies. For a complete table of the evaluated models see supplementary material Table S1. For simplicity, we choose the model with dummies to make the predictions.
225C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
8
13, 2020. Apple also uses people’s mobile devices to track their location (monitoring the
requests made to the Apple map application). For Apple, we use the mobility measure
based on driving. Data is available for the period January 13 to June 30.2 For purposes of
comparability between the Google and Apple datasets, we change the baseline to
February 17. We thus obtain a dataset for the period February 15 to June 30, with each
row including two columns: the percent change of total POS expenditure from state s in
week w, and the mean percent change in mobility from each source from state s in week
w. The percent change is with respect to February 17 in all columns.
∆%𝑃𝑂𝑆 , 𝛽∆%𝑀𝑜𝑏𝑖𝑙𝑖𝑡𝑦 , 𝛿 𝛿 𝜀 , (3)
The regression controls for fixed effects of week and state. The first control is for shocks
that affect all states at the same time, and the second is for permanent differences across
states. For example, some states may specialize in occupations or industries that make
them either more resilient or more susceptible to an economic shock, and this
specialization may at the same time be correlated with geographic mobility.3
3. DESCRIPTIVE RESULTS
The first case of COVID-19 in Mexico was diagnosed on February 27, later than
in European countries. On March 14, the government announced the suspension of non-
essential activities and rescheduled mass events. A soft lockdown began on March 23.
The government has taken different steps to address the health and economic shocks.
First, it implemented recommendations for social distancing, travel restrictions, and the
suspension of non-essential activities to prevent the spread of the virus. Second, the
2 Two days, May 11 and May 12, were not available. We impute values for these days with the mean values for May 10 and May 13. 3 In particular, some states may be more prepared for telecommuting than others, making them more resistant to employment losses. If the latter states show greater mobility and expenditure, that could bias the elasticity estimate.
226C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
9
government and the Banco de México have taken action to mitigate the effects of the
pandemic. Like other central banks across the world, the Banco de México has
implemented measures to provide liquidity to the market, injecting the equivalent of 3.3
percent of GDP into the economy.4 The fiscal policy response has been more limited: it
has offered access to microcredits and has implemented a frontloaded payment of some
social programs (close to 1 percent of GDP). The government has also announced an
austerity program and the continuation of some public works.
Table 1 shows the main descriptive statistics for May 2019 and May 2020,
including the total amount spent in POS terminals, the average amount of each
transaction, and the share of expenditures in each group. For simplicity the data is grouped
into 12 categories: tourism (travel agencies and hotels), education (universities, colleges,
basic education, and daycare), health care (pharmacies, hospitals, physicians, and
dentists), food services (restaurants and fast food), trade (wholesale and retail),
transportation (air transportation, ground transportation, tolls, parking lots, and car
rental), insurance, telecommunications, supermarkets, big-box stores, and others.
The average transaction amount did not change substantially. It was $601 in May
2019 and $589 in May 2020 (in constant MXN pesos of July 2018). However, there is an
overall decline of approximately $34 billion, or 16 percent, representing an average
monthly decline in total private consumption of 2.5 percent a month. The sectors with the
largest expenditures in 2019 (a combined total of 80 percent) were big-box stores, trade,
gasoline, food services, and other. In May 2020, most sectors showed reduced total POS
transactions. Services related to tourism, food services, and transportation were hit
4 These measures have included bond swaps, loosened rules for minimum deposits from commercial banks, and facilities to swap assets with the central bank in order to obtain credit. These measures have the goal of directing credit to small and medium-sized business.
227C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
10
especially hard. However, insurance, telecommunications, big-box stores, and other
maintained or increased sales.
Table 1. Descriptive Statistics
May 2019 May 2020
Total Amount (millions of pesos)
Avg. Transaction
amount (pesos)
Share (%)
Total Amount (millions of pesos)
Avg. Transaction
amount (pesos)
Share (%)
Total $206,669 $ 601 $172,800 $589
Tourism $5,333 $ 2,580 2.6 $786 $1,988 0.5
Education $6,851 $ 4,026 3.3 $4,607 $4,639 2.7
Health Care $8,239 $ 524 4.0 $7,757 $485 4.5
Food Services $11,747 $ 383 5.7 $2,681 $257 1.6
Trade $42,312 $ 473 20.5 $32,978 $462 19.1
Transportation $8,961 $ 589 4.3 $1,833 $286 1.1
Insurance $5,153 $1,811 2.5 $5,445 $2,207 3.2
Telecomm. Services $6,394 $ 701 3.1 $6,779 $525 3.9
Gasoline $18,400 $616 8.9 $ 10,979 $538 6.4
Other $35,244 $641 17.1 $ 41,564 $601 24.1
Supermarkets $28 $348 0.0 $19 $314 0.0
Big-Box Stores $58,007 $630 28.1 $57,373 $692 33.2
Notes: Authors’ calculations. Amounts are in constant MXN for July 2018.
3.1 Aggregate Results
Figure 1 shows smoothed lines of daily expenditure in POS terminals for 2019
and 2020. For comparison purposes, the series are in relative terms with respect to January
14 of each year. The red line is the index for 2019 and the blue line for 2020. Using the
method described above, we obtain a prediction for 2020 using data for the early part of
the year. The green line is the prediction for 2020. Before the lockdown, the patterns for
2019 and 2020 are similar. When the lockdown starts, POS expenditures fall drastically.
The worst days were in mid-April, with expenditures about 35 percent lower than in 2019
228C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
11
or in the prediction for 2020. After that point, expenditures slowly started to recover. By
late May and early June, the shortfall was only about 15 percent lower than the prediction.
Figure 1. Effect of COVID-19 on expenditures in POS terminals
Notes: Authors’ calculations. Lines are smoothed using a moving average for the previous two weeks. Predicted line is obtained with equation (3), an OLS of the amount in 2020 with the amount in 2019. Expenditures are in constant pesos (MXN) of July 2018. Expenditures in January 2020 are 9 percent larger than in January 2019.
Figure 2 shows the decline in POS expenditures by month (constant pesos of July
2018), with comparisons to 2019 and the predicted expenditures for 2020. The greatest
decline is in April, with expenditures 30 percent lower than predicted and 23 percent
lower than in the corresponding period in 2019. Subsequent months show lesser declines:
May is 22 percent and June is 18 percent below the prediction. The total decline in POS
expenditures from the predicted figure for April through June is around $149 billion
MXN, a loss of 3.9 percent of an average quarter of private consumption in 2019, and a
loss of 2.6 percent of an average quarter of GDP.
229C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
12
Figure 2. Decline in POS expenditures
A. Absolute decline B. Percent decline
Note: Authors’ calculations. This graph shows the difference between actual and predicted values, and the difference between actual 2020 and 2019 values (in constant pesos of July 2018).
3.2 Results by Sector
Figure 3 shows the change in consumption patterns by sector. The lines are
smoothed using a moving average of the previous two weeks (Leatherby and Gelles
2020), and the comparison is to the predicted sales in each sector. The comparison with
respect to 2019 can be found in the Supplementary Materials. After the beginning of the
lockdown, there is a sharp decline in education, tourism, food services, and transportation.
Only education recovers, but at the end of May it is still about 40 percent below the
prediction. Tourism, food services, and transportation fall from 80 to 90 percent by mid-
April. Because of the decline in mobility and in domestic prices, POS expenditures for
gasoline decrease almost 50 percent in mid-April, and by the end of May they were still
approximately 35 percent below the prediction. Expenditures in June have been relatively
stable, with a slight recovery in most cases.
Similar to the experience of other countries, POS expenditures in big-box stores
increased in the last two weeks of March, an effect of panic buying to stockpile goods.
230C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
13
Other sectors, like insurance, health care, and telecommunications, were not affected by
the mobility restrictions. At least with insurance and telecommunications, this is likely
related to direct billing options as well as the inelasticity of demand for this type of goods.
While in the U.S. there was a large decline in health expenditures in April (Chetty et al.
2020), in Mexico the decline was smaller and it quickly recovered, by the end of May.
Figure 3. Changes in consumption patterns by sector. Smoothed lines.
Notes: Authors’ calculations. Comparison is to predicted sales in each sector. Constant pesos (MXN) of July 2018. Smoothed with moving average of the previous two weeks.
Figure 4 summarizes previous estimates. It indicates the percent difference of POS
expenditures in 2020 with respect to the predicted expenditures and with respect to the
same period in 2019 (in constant pesos). The losses total 23 percent of predicted
expenditures: one quarter of expected POS sales did not take place. Total expenditures
were 18 percent lower than in the same period in 2019. Comparisons are difficult because
231C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
14
lockdowns were implemented at different times in different countries, but the Mexican
loss estimate is among the lowest. In the last two weeks of March, France and Spain had
expenditure losses of 50 percent (Bounie et al. 2020; Carvalho et al. 2020); in April,
Portugal had losses of 55 percent (Carvalho, Peralta, and Pereira, 2020) and Denmark had
more moderate losses of approximately 30 percent.
Figure 4. Summary of expenditure losses. April-June 2020.
Note: Authors’ calculations. This graph shows the change in expenditures relative to 2019 values and to predicted values for 2020. Constant pesos (MXN) of July 2018.
The comparison with the U.S. depends on the source. The estimates of Baker et
al. (2020) imply a decline of 50 percent, while Chetty et al. (2020) find a decline of 30
percent in the last two weeks of March. In fact, the change in All Expenditures in Figure
3 is very close to that found in Chetty et al. (2020) (see Figure S3 in Supplementary
Materials). The decline in expenditures in the U.S. was larger before mid-April. Stimulus
payments began on April 15 in the U.S., and POS expenditures recovered faster around
232C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
15
that date. The decline in POS expenditures from January to mid-June was 10 percent in
the U.S., while in Mexico it was still 20 percent. There is significant heterogeneity across
sectors, however. Those affected most severely in Mexico were tourism, food services,
and transportation, where expenditures declined approximately 80 percent. This is similar
to what previous studies have found in Denmark, Spain, the United States, and other
countries (Andersen et al. 2020a; Baker et al. 2020; Carvalho et al. 2020; Chetty et al.
2020; Leatherby and Gelles 2020).
Some sectors in Mexico even had gains or only small losses. Expenditures on
insurance increased slightly in the period, and expenditures on telecommunications
decreased slightly. We interpret these sectors as supplying highly inelastic necessities.
Expenditures in big-box stores decreased by 5.4 percent. The pattern for these stores is
mixed: in mid-March their expenditures increased, in mid-April they declined, and by the
end of May they recovered. This group includes large supermarkets (such as Walmart and
Soriana) as well as department stores (such as Liverpool, Palacio de Hierro, and Sears).
It is likely that sales increased in large supermarkets and decreased in department stores.
There were decreased sales in health care, gasoline, trade, small supermarkets, and
other, which accounted for close to 50 percent of all expenditures in 2019 (Table 1). The
decline in trade and gasoline (28 and 38 percent, respectively) is directly related to
restrictions in mobility.
3.3 Results by State
We estimate the model in equation (1) for each state in Mexico. Figure 5 shows
percent losses by state with respect to the predictions of the model. States shown in purple
are the hardest hit and those in yellow are the least affected. The hardest hit regions
depend on international tourism: Quintana Roo and Yucatan in the south, as well as
233C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
16
Guerrero and Nayarit. These states lost all expected revenue from the spring vacation
season. Other states closer to Mexico City are also greatly affected: Michoacán, Estado
de México, Puebla, and Morelos, probably related to the loss of domestic tourism around
Easter. Mexico City is not as affected as other states. We suspect that here the effects of
the pandemic were partially compensated by online sales, but our data unfortunately does
not distinguish online from other sales. Finally, states in the north are not as affected as
the rest, an effect of greater mobility than in the rest of Mexico, as explained in the next
section.
Figure 5. Losses by state in total POS expenditure (percent).
Notes: Authors’ calculation. The map shows the percent change of POS expenditures from April to June with respect to the predicted sales for each state. Constant pesos of July 2018
234C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
17
4. CONSUMPTION AND MOBILITY
We use geographic mobility data from Google (2020) and Apple (2020) through
June 2020. There is an ongoing debate about the relationship between mobility and POS
expenditures. The case and evidence from Sweden are relevant.5 Unlike other European
countries, Sweden did not impose a lockdown in response to the COVID-19 pandemic,
which was responsible for a higher mortality rate than in similar Nordic countries. One
might expect that the lack of restrictions on mobility at least lessened the economic effects
of the pandemic. However, Andersen et al. (2020b) found that this was not the case.
Sweden experienced a 25 percent reduction in POS expenditures from March 11 to April
12; the corresponding figure for Denmark was 29 percent. Apple’s measure of driving
mobility for early April shows a reduction in Sweden of 12 percent and a reduction in
Denmark of 32.4 percent. The between-country variation suggests that the elasticity of
mobility is around 0.20. However, elasticity may depend on the relative importance of
internet sales, which depends in turn on the depth of the financial sector. In a less
developed country like Mexico, in-person sales and therefore mobility may matter much
more than in developed economies.
To show how mobility and expenditures are related, we use Google’s measure of
workplace mobility and Apple’s measure of driving mobility. We calculate mobility and
expenditure patterns for each of the 20 weeks and 32 states under study. We thus have a
panel dataset with 640 observations. The patterns are shown in Figure 6. The variation in
the mobility measures is positively correlated with the variation in the total amount spent.
Panel A uses Google’s workplace mobility and it finds a coefficient of 0.7 using a simple
OLS regression. Panel B uses Apple’s driving mobility and it finds a coefficient of 0.9
5 https://www.politico.eu/article/swedens-cant-escape-economic-hit-with-covid-19-light-touch/, https://www.ft.com/content/93105160-dcb4-4721-9e58-a7b262cd4b6e.
235C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
18
with the same type of regression. States with the largest declines in mobility are related
to the largest declines in expenditures at the weekly level.
Figure 6. Relationship between mobility (Google and Apple)
and POS expenditures.
A. Google: Workplace Mobility B. Apple: Driving Mobility
Notes: Authors’ calculations. Each dot is the percent change of mobility or POS expenditures (constant pesos of July 2018) in week w with respect to February 17 for each of the 32 states in Mexico. Period of estimation is February 15 to June 30.
In order to analyze this claim more carefully, we estimate different versions of
equation (3). Panel A in Table 2 estimates the relationship between changes in POS
expenditures and mobility in Mexico including week and state fixed effects. These effects
control for permanent differences across states (for example, density or geographic
characteristics) as well as for temporal shocks that affect all states at the same time. Table
2 shows the results for all expenditures as well as for expenditures differentiated by credit
versus debit card. The elasticity coefficient for total expenditure using Google’s mobility
is 0.73; for Apple’s mobility it is 0.93. These estimates, which exploit the within-state
variation, are very similar to those obtained simply by pooling the spending and mobility
information (Figure 6). All of the estimates in Panel A are very precisely estimated and
they are all statistically significant. The elasticity using Apple’s mobility information is
not statistically different from 1. The elasticity for credit card spending is greater than for
debit cards, regardless of the mobility indicator used.
236C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
19
Panel B in Table 2 estimates the same spending-mobility relationship but instead
of including fixed effects, it includes as an additional control variable the proportion of
work that can be performed by telecommuting from home in each Mexican state, as
estimated by Monroy-Gómez-Franco (2020). This specification exploits both the between
and the within variation across states in Mexico to estimate the elasticity of POS
expenditures with respect to mobility. The elasticity results obtained with this
specification have larger standard errors, but they are similar for both mobility indicators
(0.91). In both cases, they are statistically different from 0 but not from 1. As before,
credit cards are more elastic with respect to mobility than debit cards. These estimated
elasticities are also much larger than that implied by Andersen et al. (2020b) for the case
of Sweden (0.2).
Table 2. Elasticity Estimates: Change in % POS Expenditures with Respect to
Change in % Mobility
Google: Workplace Mobility Apple: Driving Mobility
Total Credit Debit Total Credit Debit
A. Including state fixed effects Coefficient 0.73 1.09 0.56 0.93 1.33 0.74
Standard Error [0.05] [0.05] [0.04] [0.06] [0.07] [0.05]
R2 0.45 0.54 0.35 0.46 0.44 0.41
Total Obs. 640 640 640 640 640 640
B. Controlling for telecommuting (without state fixed effects)
Coefficient 0.91 1.34 0.67 0.91 1.03 0.85
Standard Error [0.45] [0.51] [0.42] [0.20] [0.29] [0.18]
R2 0.57 0.63 0.51 0.65 0.66 0.62
Total Obs. 640 640 640 640 640 640
Notes: Authors’ calculations. The dependent variable is the percent change in POS expenditures in week w with respect to February 17 for each state in Mexico, and the independent variable is the percent change in mobility for the same period. The regression in Panel A includes fixed effects for state and week. Estimation period is February 15 to June 27. Panel B includes dummies for weeks and proportion of telecommuting (defined as in Monroy-Gómez-Franco 2020). Standard errors clustered at the state level in brackets.
To further analyze these results, in Figure 7 we show Apple’s mobility measure
(blue line) and POS expenditures (red line) in high- and low-mobility states in Mexico.
237C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
20
In high-mobility states, mobility and POS expenditures declined close to 10 percent from
early in the year to the end of May. By mid-June, mobility and expenditures in these states
were similar to pre-pandemic levels. In low-mobility states the decline was close to 25
percent in mid-May and by mid-June it was still around 15 percent below pre-pandemic
levels. The high correlation between spending and mobility in both types of states is
evident. As expected, the estimates of the elasticity of spending with respect to mobility
are also very high for each of these groups of states: around 0.80 for high-mobility states
and 1.04 for low-mobility states, and in both groups the elasticity for credit cards is larger
than for debit cards. These elasticity estimates are between four and five times the implied
elasticity estimated by Andersen et al. (2020b).
Figure 7. Apple’s mobility and POS expenditures in high- versus low-mobility
states.
A. High mobility B. Low mobility.
Notes: Authors’ calculations. High-mobility states include Aguascalientes, Campeche, Chihuahua, Coahuila, Colima, Durango, Guerrero, Michoacán, Morelos, San Luis Potosí, Sinaloa, Sonora, Tamaulipas, Tlaxcala, Veracruz, and Zacatecas. Low-mobility states include Baja California, Baja California Sur, Chiapas, Mexico City, Estado de México, Guanajuato, Hidalgo, Jalisco, Nayarit, Nuevo León, Oaxaca, Puebla, Querétaro, Quintana Roo, Tabasco, and Yucatán. Mobility refers to driving mobility measured by Apple.
Why is the elasticity of POS expenditures to mobility larger in Mexico? We
conjecture that this difference is driven mainly by the strength (or lack thereof) of e-
commerce, financial inclusion, and internet penetration. As mentioned in the introduction,
238C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
21
financial inclusion is lower in Mexico than in China, the United States, and European
countries. In 2014, the proportion of individuals in Mexico with an account at a financial
institution was 40 percent, while it is 80 percent in China and close to 100 percent in
developed countries. If we consider that internet penetration is lower as well, then we
have a weaker market for e-commerce in Mexico than elsewhere. Indeed, results from the
United Nations Conference on Trade and Development (2016) show that Mexico has an
e-commerce readiness index much lower than other countries (Mexico’s index is 49.1
while the U.S.’s is 82.6). POS transactions thus depend much more on mobility in Mexico
than in other countries.
Finally, we cannot test Chetty’s et al. (2020) claim that the channels of the decline
in POS expenditures are mainly rich individuals in fear of contagion. We attempt to
compare our results, at least at the aggregate level, by computing POS expenditures by
credit versus debit cards, which are highly segregated in Mexico. We calculate that
approximately 77 percent of credit cards and 62 percent of debit cards are held by
individuals in the top 30 percent of the wealth distribution (see figures in Supplementary
Materials). The decline in POS expenditures is larger for credit cards (28.6 percent) than
for debit cards (10 percent). The elasticity of POS expenditures with respect to mobility
is also much larger for credit cards than for debit cards. We thus conjecture that the
decline in POS expenditures is partially driven by richer individuals concerned for their
health, as in Chetty et al. (2020).
5. SUMMARY
This paper analyzes consumption patterns in Mexico using the universe of POS
transactions for the period from January 2019 to June 2020. Unlike some other countries,
Mexico implemented a soft lockdown as well as a moderate countercyclical fiscal policy.
239C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
22
We find that POS expenditures for the April-June quarter are 23 percent less than they
would have been in the absence of the pandemic. This difference is less than that
calculated for European countries using similar data, and comparable to that reported for
the U.S. by Chetty et al. (2020), also using results based on live POS data.
The losses we find for Mexico are heterogeneous across economic sectors and
region. As in other studies, the more severely affected sectors are those related to tourism
(travel agencies and hotels), food services (such as restaurants), and transportation. States
that benefit more directly from tourism (beach resorts and other tourist destinations) were
also more affected.
There is a debate about whether mobility patterns affect POS expenditures and
thus economic activity. We find that the elasticity of POS expenditures with respect to
mobility is close to 1 (0.93 using Apple’s measure of mobility in one specification and
0.91 for both Google’s and Apple’s measures of mobility in another). These estimates are
much larger than the implied elasticity estimated by Andersen et al. (2020b) for Sweden.
Our estimate likely indicates that POS expenditures in developing countries with
shallower financial sectors are more responsive to mobility patterns than developed
countries. It also suggests that mobility indicators, which can be observed almost in real
time, could serve as a good proxy for the behavior of expenditures in some economies.
240C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
23
REFERENCES
Andersen, Asger, Emil Toft Hansen, Niels Johannesen and Adam Sheridan
(2020a). “Consumer Responses to the COVID-19 Crisis: Evidence from Bank Account
Transaction Data". Centre for Economic Policy Research.
http://www.cepr.org/active/publications/discussion_papers/dp.php?dpno=14809
Andersen, Asger, Emil Toft Hansen, Niels Johannesen, and Adam Sheridan
(2020b). “Pandemic, Shutdown and Consumer Spending: Lessons from Scandinavian
Policy Responses to COVID-19”. Working Paper arXiv:2005.04630v1.
Apple Inc. (2020). "Apple Mobility Trends reports".
https://www.apple.com/covid19/mobility/ (Accessed: June 5th, 2020)
Baker, Scott R., R.A. Farrokhnia, Steffen Meyer, Michaela Pagel and
Constantine Yannelis (2020). “How Does Household Spending Respond to an
Epidemic? Consumption During the 2020 COVID-19 Pandemic”. Covid Economics 18,
15 May 2020: 73-108.
Bounie, David, Youssouf Camara and John W. Galbraith (2020). “Consumers'
Mobility, Expenditure and Online-Offline Substitution Response to COVID-19:
Evidence from French Transaction Data”. CIRANO Working Papers 2020-28.
Carvalho, Bruno P., Susana Peralta and Joao Pereira (2020). “What and How
did People Buy during the Great Lockdown? Evidence from Electronic Payments”. Covid
Economics 28, 12 June 2020: 119-158.
Carvalho, Vasco, M., Stephen Hansen, Álvaro Ortiz, Juan Ramón García,
Tomasa Rodrigo, Sevi Rodríguez Mora, and José Ruiz (2020). "Tracking the COVID-
19 Crisis with High-Resolution Transaction Data". Centre for Economic Policy Research.
https://cepr.org/active/publications/discussion_papers/dp.php?dpno=14642
241C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
24
Chen, Haiqiang, Wenlan Qian and Qiang Wen (2020). “The Impact of the
COVID-19 Pandemic on Consumption: Learning from High Frequency Transaction
Data”. SSRN. http://dx.doi.org/10.2139/ssrn.3568574
Chetty, Raj, John N. Friedman, Nathaniel Hendren, Michael Stepner, and
the Opportunity Insights Team (2020). “How Did COVID-19 and Stabilization
Policies Affect Spending and Employment? A New Real-Time Economic Tracker Based
on Private Sector Data.” National Bureau of Economic Research, NBER Working Paper
27431, https://doi.org/10.3386/w27431.
Google LLC. (2020). "Google COVID-19 Community Mobility Reports".
https://www.google.com/covid19/mobility/ (Accessed: June 5th, 2020)
Guerrieri, Veronica, Guido Lorenzoni, Ludwig Straub, and Ivan Werning
(2020). “Macroeconomic implications of COVID-19: Can negative supply shocks cause
demand shortages?” National Bureau of Economic Research, NBER Working Paper
26918, https://doi.org/10.3386/w26918.
Hacioglu, Sinem, Diego Känzig and Paolo Surico (2020). "Consumption in the
time of Covid 19: Evidence from UK transaction data". Centre for Economic Policy
Research, CEPR Discussion Paper 14733.
http://www.cepr.org/active/publications/discussion_papers/dp.php?dpno=14733
Instituto Nacional de Estadística y Geografía. (2018). “Encuesta Nacional de
Inclusión Financiera (ENIF) 2018”. https://www.inegi.org.mx/programas/enif/2018/
(Accessed: June 5th, 2020)
Kahn, Lisa B., Fabian Lange, and David G. Wiczer (2020). “Labor Demand in
the Time of COVID-19: Evidence from Vacancy Postings and UI Claims.” National
Bureau of Economic Research, NBER Working Paper 27061,
https://doi.org/10.1017/CBO9781107415324.004.
242C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
25
Lewis, Daniel, Karel Mertens, and James H. Stock (2020). “US economic
activity during the early weeks of the SARS-Cov-2 outbreak”. National Bureau of
Economic Research, NBER Working Paper 26954, https://doi.org/10.3386/w26954
Leatherby, Lauren and David Gelles (2020). “How the Virus Transformed the
Way Americans Spend Their Money” New York Times. April 11.
https://www.nytimes.com/interactive/2020/04/11/business/economy/coronavirus-us-
economy-spending.html?referringSource=articleShare
Monroy-Gómez-Franco, Luis (2020). “¿Quién puede trabajar desde casa?
Evidencia desde México.”, Documento de Trabajo no. 06/2020. Centro de Estudios
Espinosa Yglesias. https://ceey.org.mx/wp-content/uploads/2020/05/06-Monroy-
G%C3%B3mez-Franco-2020.pdf
United Nations Conference on Trade and Development (2016). UNCTAD B2B
E-Commerce Index 2016 (Vol. 7).
https://unctad.org/en/PublicationsLibrary/tn_unctad_ict4d07_en.pdf
World Bank (2020). Data indicators. Data.worldbank.org. Series:
https://data.worldbank.org/indicator/FD.AST.PRVT.GD.ZS?locations=US-ES-GB-CN-
DK-FR&name_desc=true,
https://databank.worldbank.org/reports.aspx?source=1251&series=GPSS_4
243C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
26
Supplementary Materials
Table S1. RMSPE Tested Models
RMPSE February 19-March 11 Model # Model description Total Credit Debit
1 OLS 547.83 267.93 414.86 2 OLS dummies 425.40 197.64 299.30 3 ARIMA(0, 0 ,0) 547.83 267.93 414.86 4 ARIMA(0, 1 ,0) 598.36 268.81 479.59 5 ARIMA(0, 2 ,0) 6153.22 2450.88 1082.39 6 ARIMA(0, 0 ,1) 551.20 259.99 413.24 7 ARIMA(0, 1 ,1) 538.81 281.95 396.49 8 ARIMA(0, 2 ,1) 695.04 450.37 434.59 9 ARIMA(0, 0 ,2) 559.96 278.89 420.63
10 ARIMA(0, 1 ,2) 544.82 271.71 391.56 11 ARIMA(0, 2 ,2) 549.44 277.89 412.70 12 ARIMA(1, 0 ,0) 547.46 267.40 413.63 13 ARIMA(1, 1 ,0) 519.04 296.20 455.55 14 ARIMA(1, 2 ,0) 2892.21 645.85 817.03 15 ARIMA(1, 0 ,1) 561.43 282.26 414.88 16 ARIMA(1, 1 ,1) 539.55 280.94 394.15 17 ARIMA(1, 2 ,1) 586.35 386.94 451.30 18 ARIMA(1, 0 ,2) 556.11 276.08 418.16 19 ARIMA(1, 1 ,2) 531.36 272.46 394.61 20 ARIMA(1, 2 ,2) 548.14 276.24 433.78 21 ARIMA(2, 0 ,0) 543.16 276.84 420.12 22 ARIMA(2, 1 ,0) 509.93 253.88 440.78 23 ARIMA(2, 2 ,0) 1937.40 963.75 2354.00 24 ARIMA(2, 0 ,1) 545.08 265.54 419.62 25 ARIMA(2, 1 ,1) 537.28 285.54 403.67 26 ARIMA(2, 2 ,1) 512.18 246.57 436.00 27 ARIMA(2, 0 ,2) 436.47 277.93 422.59 28 ARIMA(2, 1 ,2) 536.05 273.96 399.86 29 ARIMA(2, 2 ,2) 548.95 288.44 424.83 30 ARIMA(0, 0 ,0) dummies 425.40 197.64 299.30 31 ARIMA(0, 1 ,0) dummies 476.93 225.70 276.60 32 ARIMA(0, 2 ,0) dummies 7886.29 2126.09 1792.76 33 ARIMA(0, 0 ,1) dummies 456.05 218.66 296.66 34 ARIMA(0, 1 ,1) dummies 415.38 202.44 284.85 35 ARIMA(0, 2 ,1) dummies 798.20 557.02 318.68 36 ARIMA(0, 0 ,2) dummies 477.72 241.28 306.00 37 ARIMA(0, 1 ,2) dummies 425.18 219.22 282.93 38 ARIMA(0, 2 ,2) dummies 405.73 650.48 339.80 39 ARIMA(1, 0 ,0) dummies 425.47 196.96 297.46 40 ARIMA(1, 1 ,0) dummies 417.76 225.53 257.01 41 ARIMA(1, 2 ,0) dummies 4905.65 980.97 2054.57 42 ARIMA(1, 0 ,1) dummies 469.24 212.66 310.32 43 ARIMA(1, 1 ,1) dummies 415.31 199.39 283.38 44 ARIMA(1, 2 ,1) dummies 719.47 593.73 298.22 45 ARIMA(1, 0 ,2) dummies 476.83 211.87 308.88 46 ARIMA(1, 1 ,2) dummies 429.18 214.44 286.64 47 ARIMA(1, 2 ,2) dummies 714.69 611.71 265.89 48 ARIMA(2, 0 ,0) dummies 438.86 269.24 303.32 49 ARIMA(2, 1 ,0) dummies 456.70 248.57 259.42
244C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
27
50 ARIMA(2, 2 ,0) dummies 827.67 1036.05 1455.96 51 ARIMA(2, 0 ,1) dummies 455.71 265.13 313.65 52 ARIMA(2, 1 ,1) dummies 420.19 251.79 288.18 53 ARIMA(2, 2 ,1) dummies 438.43 250.41 279.31 54 ARIMA(2, 0 ,2) dummies No convergence 260.20 311.00 55 ARIMA(2, 1 ,2) dummies 415.41 268.92 288.83 56 ARIMA(2, 2 ,2) dummies 381.85 278.92 270.50
Notes: Models estimated for first 7 weeks of 2020. All the models are controlled for the amount in the corresponding weeks of 2019. Dummies refers to paydays, Mondays, Fridays, and December. RMPSE calculated for February 19 to March 11.
Figure S1. Changes in consumption patterns by sector with respect to
corresponding period in 2019.
Notes: Authors’ calculations. Comparison is to corresponding period in 2019 (in constant pesos of July
2018).
245C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
28
Figure S2. Changes in consumption patterns by sector with respect to
corresponding period in 2019.
A. Credit cards
B. Debit cards
Notes: Authors’ calculations. Comparison is to corresponding period in 2019 (in constant pesos of July 2018).
246C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
29
Figure S3. Comparison of expenditures in Mexico and U.S. Smoothed lines.
Note: Authors’ calculations for the Mexican series. For the U.S., we used the calculation published by Chetty et al. (2020) and the https://tracktherecovery.org/ webpage. Both series use exactly the same construction. We first take a seven-day moving average, then we divide the 2020 series by the 2019 calendar day-month values. Finally, we divide the series by its average value for January 4-31.
247C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
30
Figure S4. Losses by state in total POS expenditure.
Notes: Authors’ calculation. The map shows the percent change in POS expenditures from April to June with respect to sales for each state in the same period in 2019. Constant pesos of July 2018.
Table S3. Elasticity Estimates: Change in Percent Sales with Respect to Change in Percent Mobility, Panel by Day
Google: Workplace Mobility Apple: Driving Mobility
Total Credit Debit Total Credit Debit Coefficient 0.59 0.83 0.49 0.84 1.17 0.68 Standard
Error [0.04] [0.05] [0.03] [0.05] [0.08] [0.04]
R2 0.27 0.31 0.21 0.31 0.32 0.26
Total Obs. 4384 4384 4384 4384 4384 4384 Notes: Authors’ calculations. The dependent variable is the percent change in POS expenditures on day d with respect to February 17 for each state in Mexico; the independent variable is the percent change in mobility. The regression includes fixed effects for state and day. The estimation period is February 15 to June 30. Clustered standard errors at state level.
248C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
31
Table S4. Elasticity estimates: Change in % POS expenditures with respect to change in % Mobility
Apple: High Driving Mobility Apple: Low Driving Mobility
Total Credit Debit Total Credit Debit Coefficient 0.80 1.21 0.63 1.04 1.43 0.83 Standard
Error [0.05] [0.08] [0.04] [0.09] [0.11] [0.07]
R2 0.31 0.35 0.24 0.57 0.56 0.52
Total Obs. 320 320 320 320 320 320 Notes: Authors’ calculations. The dependent variable is the percent change in POS expenditures in week w with respect to February 17 for each state in Mexico; the independent variable is the percent change in mobility for the same period. The regression includes fixed effects for state and week. Estimation period is February 15 to June 27. Clustered standard errors at the state level in brackets. High mobility states include Aguascalientes, Campeche, Chiapas, Chihuahua, Coahuila, Colima, Durango, Michoacán, Morelos, San Luis Potosi, Sinaloa, Sonora, Tamaulipas, Tlaxcala, Veracruz, and Zacatecas; low mobility states include Baja California, Baja California Sur, Mexico City, Estado de México, Guanajuato, Guerrero, Hidalgo, Jalisco, Nayarit, Nuevo León, Oaxaca, Puebla, Querétaro, Quintana Roo, Tabasco, and Yucatán. High mobility refers to driving mobility measured by Apple.
249C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
32
Figure S5. Cardholding by household asset index decile
Notes: Authors’ calculation with data from the INEGI intergenerational social mobility module (2016). Household asset index was constructed with PCA of the covariance matrix using asset holding (television, vehicles, home ownership, telephone, internet access, radio, DVD, blender, toaster, microwave, refrigerator, stove, washing machine, iron, sewing machine, fan, tablet computer, videogame console, computer, printer, and livestock) and years of schooling.
9.4 1
0
23
16
27
39
31
58
65
75
2.1 2.5
8.5
3.9
7.4
16
13
32
39
54
0
20
40
60
80%
1 2 3 4 5 6 7 8 9 10Household Asset Index Decile
Population with debit card Population with credit card
250C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
33
Figure S6. Cardholding by household asset index position A. Debit card B. Credit card
Notes: Authors’ calculation with data from INEGI intergenerational social mobility module (2016). The graph shows the percent of credit (debit) cardholding by household asset index distribution. Household asset index was constructed with PCA of the covariance matrix using asset holding (television, vehicles, home ownership, telephone, internet access, radio, DVD, blender, toaster, microwave, refrigerator, stove, washing machine, iron, sewing machine, fan, tablet computer, videogame console, computer, printer, and livestock) and years of schooling.
38%
62%
Rest of distribution Top 30%
23%
77%
Rest of distribution Top 30%
251C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252
COVID ECONOMICS VETTED AND REAL-TIME PAPERS
34
Figure S7. Comparison of mobility between United States and Mexico
Notes: Authors’ calculation. Mobility uses Apple (2020) driving mobility. Base index is January 2020.
252C
ovid
Eco
nom
ics 3
8, 1
6 Ju
ly 2
020:
218
-252